Cognitive Radio Technology
This page intentionally left blank
Cognitive Radio Technology
Edited by Bruce A. Fette
AMSTERDAM
•
•
•
NEW YORK
BOSTON
•
SAN FRANCISCO
•
OXFORD
•
HEIDELBERG
•
PARIS
SINGAPORE
•
•
•
LONDON
SAN DIEGO
SYDNEY
Newness is an important of Elsevier
•
TOKYO
Newnes is an imprint of Elsevier
30 Corporate Drive, Suite 400, Burlington, MA 01803, USA
Linacre House, Jordan Hill, Oxford OX2 8DP, UK
Copyright © 2006, Elsevier Inc. All rights reserved.
No part of this publication may be reproduced, stored in a retrieval system, or
transmitted in any form or by any means, electronic, mechanical, photocopying,
recording, or otherwise, without the prior written permission of the publisher.
Permissions may be sought directly from Elsevier’s Science & Technology Rights
Department in Oxford, UK: phone: (44) 1865 843830, fax: (44) 1865 853333,
E-mail: HYPERLINK "mailto:[email protected]" [email protected].
You may also complete your request on-line via the Elsevier homepage (http://elsevier.com),
by selecting “Support & Contact” then “Copyright and Permission” and then “Obtaining
Permissions.”
Recognizing the importance of preserving what has been written, Elsevier prints its books
on acid-free paper whenever possible.
Library of Congress Cataloging-in-Publication Data
Cognitive radio technology / edited by Bruce A. Fette.—1st ed.
p. cm.—(Communications engineering series)
Includes bibliographical references and index
ISBN-13: 978-0-7506-7952-7 (alk. paper)
ISBN-10: 0-7506-7952-2 (alk. paper)
1. Software radio. 2. Artificial intelligence. 3. Wireless communication systems. I. Fette, Bruce A.
II. Series.
TK5103.4875.C64
621.384—dc22
2006
British Library Cataloguing-in-Publication Data
A catalogue record for this book is available from the British Library.
ISBN 13: 978-0-7506-7952-7
ISBN 10: 0-7506-7952-2
For information on all Newnes publications visit our Web site at
www.books.elsevier.com
06 07 08 09 10
10 9 8 7 6 5 4 3 2 1
Typeset by Charon Tec Ltd, Chennai, India
www.charontec.com
Printed in the United States of America
2006016824
Contents
List of Contributors .................................................................. xvii
Foreword.................................................................................. xxi
Chapter 1: History and Background of Cognitive Radio
Technology Bruce A. Fette ....................................................... 1
1.1
1.2
1.3
1.4
1.5
1.6
1.7
1.8
The Vision of Cognitive Radio..............................................................1
History and Background Leading to Cognitive Radio ..........................2
A Brief History of SDR ........................................................................4
Basic SDR .............................................................................................8
1.4.1 The Hardware Architecture of an SDR..................................8
1.4.2 Computational Processing Resources in an SDR ................11
1.4.3 The Software Architecture of an SDR .................................13
1.4.4 Java Reflection in a Cognitive Radio ...................................15
1.4.5 Smart Antennas in a Cognitive Radio..................................15
Spectrum Management .......................................................................17
1.5.1 Managing Unlicensed Spectrum ..........................................18
1.5.2 Noise Aggregation ...............................................................19
1.5.3 Aggregating Spectrum Demand and Use
of Subleasing Methods.........................................................21
1.5.4 Priority Access .....................................................................22
US Government Roles in Cognitive Radio .........................................22
1.6.1 DARPA ................................................................................22
1.6.2 FCC ......................................................................................23
1.6.3 NSF/CSTB Study.................................................................23
How Smart Is Useful? .........................................................................24
Organization of this Book ...................................................................25
v
Contents
Chapter 2: Communications Policy and Spectrum
Management Paul Kolodzy .................................................. 29
2.1 Introduction.........................................................................................29
2.2 Cognitive Radio Technology Enablers................................................30
2.3 New Opportunities in Spectrum Access .............................................33
2.3.1 Current Spectrum Access Techniques .................................33
2.3.2 Opportunistic Spectrum Access...........................................39
2.3.3 Dynamic Frequency Selection .............................................42
2.4 Policy Challenges for Cognitive Radios .............................................42
2.4.1 Dynamic Spectrum Access ..................................................43
2.4.2 Security ................................................................................46
2.4.3 Communications Policy before Cognitive Radio.................48
2.4.4 Cognitive Radio Impact on Communications Policy...........49
2.4.5 US Telecommunications Policy, Beginning with
the Titanic.............................................................................49
2.4.6 US Telecommunications Policy: Keeping Pace
with Technology...................................................................51
2.5 Telecommunications Policy and Technology Impact
on Regulation ......................................................................................53
2.5.1 Basic Geometries .................................................................53
2.5.2 Introduction of Dynamic Policies ........................................56
2.5.3 Introduction of Policy-Enabled Devices ..............................58
2.5.4 Interference Avoidance ........................................................60
2.5.5 Overarching Impact .............................................................61
2.6 Global Policy Interest in Cognitive Radios.........................................61
2.6.1 Global Interest......................................................................62
2.6.2 US Reviews of Cognitive Radios for Dynamic
Spectrum Access..................................................................62
2.7 Summary .............................................................................................69
Chapter 3: The Software Defined Radio as a Platform
for Cognitive Radio Pablo Robert and Bruce A. Fette ................. 73
3.1 Introduction.........................................................................................73
3.2 Hardware Architecture ........................................................................75
3.2.1 The Block Diagram..............................................................76
3.2.2 Baseband Processor Engines................................................82
3.2.3 Baseband Processing Deployment.......................................87
3.2.4 Multicore Systems and System-on-Chip .............................89
vi
Contents
3.3 Software Architecture .........................................................................90
3.3.1 Design Philosophies and Patterns ........................................91
3.4 SDR Development and Design ...........................................................94
3.4.1 GNURadio ...........................................................................94
3.4.2 Software Communications Architecture..............................95
3.5 Applications ......................................................................................108
3.5.1 Application Software .........................................................108
3.6 Development .....................................................................................111
3.6.1 Component Development ..................................................112
3.6.2 Waveform Development ....................................................113
3.7 Cognitive Waveform Development ...................................................114
3.8 Summary ...........................................................................................116
Chapter 4: Cognitive Radio: The Technologies
Required John Polson ....................................................... 119
4.1 Introduction.......................................................................................119
4.2 Radio Flexibility and Capability .......................................................120
4.2.1 Continuum of Radio Flexibility and Capability.................120
4.2.2 Examples of Software Capable Radios..............................121
4.2.3 Examples of Software Programmable Radios ...................126
4.2.4 Examples of SDR...............................................................126
4.3 Aware, Adaptive, and CRs ................................................................126
4.3.1 Aware Radios .....................................................................126
4.3.2 Adaptive Radios.................................................................131
4.3.3 Cognitive Radios................................................................132
4.4 Comparison of Radio Capabilities and Properties ............................133
4.5 Available Technologies for CRs........................................................133
4.5.1 Geolocation ........................................................................135
4.5.2 Spectrum Awareness/Frequency Occupancy .....................135
4.5.3 Biometrics ..........................................................................136
4.5.4 Time ...................................................................................136
4.5.5 Spatial Awareness or Situational Awareness .....................138
4.5.6 Software Technology .........................................................138
4.5.7 Spectrum Awareness and Potential for Sublease
or Borrow ...........................................................................144
4.6 Funding and Research in CRs...........................................................144
4.6.1 Cognitive Geolocation Applications..................................146
4.6.2 Dynamic Spectrum Access and Spectrum Awareness ......148
vii
Contents
4.6.3 The Rendezvous Problem ..................................................153
4.6.4 CR Authentication Applications ........................................155
4.7 Timeline for CRs...............................................................................156
4.7.1 Decisions, Directions, and Standards.................................157
4.7.2 Manufacture of New Products ...........................................157
4.8 Summary and Conclusions................................................................158
Chapter 5: Spectrum Awareness
5.1
5.2
5.3
5.4
5.5
5.6
5.7
5.8
5.9
5.10
Preston Marshall .................... 163
Introduction .....................................................................................163
The Interference Avoidance Problem..............................................164
Cognitive Radio Role ......................................................................165
Spectral Footprint Minimization .....................................................166
Creating Spectrum Awareness.........................................................168
5.5.1 Spectrum Usage Reporting ................................................168
5.5.2 Spectrum Sensing...............................................................169
5.5.3 Potential Interference Analysis ..........................................170
5.5.4 Link Rendezvous ...............................................................173
5.5.5 Distributed Sensing and Operation ....................................173
Channel Awareness and Multiple Signals in Space ........................174
Spectrally Aware Networking .........................................................176
Overlay and Underlay Techniques ..................................................178
Adaptive Spectrum Implications for Cognitive
Radio Hardware...............................................................................180
Summary: The Cognitive Radio Toolkit .........................................182
Appendix: Propagation Energy Loss...............................................183
Chapter 6: Cognitive Policy Engines
Robert J. Wellington ........... 185
6.1 The Promise of Policy Management for Radios ...............................185
6.2 Background and Definitions..............................................................185
6.3 Spectrum Policy ................................................................................187
6.3.1 Management of Spectrum Policy.......................................188
6.3.2 System Requirements for Spectrum Policy
Management.......................................................................189
6.4 Antecedents for Cognitive Policy Management................................189
6.4.1 Defense Advanced Research Projects Agency
Policy Management Projects..............................................190
6.4.2 Academic Research in Policy Management ......................191
viii
Contents
6.5
6.6
6.7
6.8
6.4.3 Commercial Applications of Policy Management .............194
6.4.4 Standardization Efforts for Policy Management................195
Policy Engine Architectures for Radio .............................................198
6.5.1 Concept for Policy Engine Operations ..............................198
6.5.2 Technical Approaches for Policy Management .................200
6.5.3 Enabling Technologies.......................................................202
Integration of Policy Engines into Cognitive Radio .........................204
6.6.1 Software Communications Architecture Integration .........204
6.6.2 Policy Engine Design.........................................................206
6.6.3 Integration of the Radio into a Network Policy
Management Architecture..................................................209
The Future of Cognitive Policy Management ...................................211
6.7.1 Military Opportunities for Cognitive Policy
Management.......................................................................211
6.7.2 Commercial Opportunities for Spectrum
Management.......................................................................212
6.7.3 Obstacles to Adoption of Policy Management
Architectures ......................................................................213
Summary ...........................................................................................214
Chapter 7: Cognitive Techniques: Physical and Link
Layers Thomas W. Rondeau and Charles W. Bostian ................ 219
7.1 Introduction.......................................................................................219
7.2 Optimizing PHY and Link Layers for Multiple-Objectives
Under Current Channel Conditions...................................................220
7.3 Defining the Cognitive Radio............................................................222
7.4 Developing Radio Controls (Knobs) and Performance
Measures (Meters).............................................................................223
7.4.1 PHY- and Link-Layer Parameters ......................................223
7.4.2 Modeling Outcome as a Primary Objective .......................227
7.5 MODM Theory and Its Application to Cognitive Radio ..................230
7.5.1 Definition of MODM and Its Basic Formulation...............230
7.5.2 Constraint Modeling ..........................................................231
7.5.3 The Pareto-Optimal Front: Finding the Nondominated
Solutions ............................................................................231
7.5.4 Why the Radio Environment Is a MODM Problem ..........232
7.5.5 GA Approach to the MODM .............................................233
ix
Contents
7.6 The Multi-objective GA for Cognitive Radios..................................239
7.6.1 Cognition Loop ..................................................................239
7.6.2 Representing Radio Parameters as Genes
in a Chromosome ...............................................................244
7.6.3 Multi-dimensional Analysis of the Chromosomes ............245
7.6.4 Relative Pooling Tournament Evaluation ..........................249
7.6.5 Example of the WSGA ......................................................249
7.7 Advanced GA Techniques ..............................................................252
7.7.1 Population Initialization.....................................................253
7.7.2 Priming the GA with Previously Observed Solutions .......254
7.7.3 CBDT Initialization of GAs...............................................255
7.8 Need for a Higher-Layer Intelligence .............................................258
7.8.1 Adjusting Parameters Autonomously to
Achieve Goals ....................................................................258
7.8.2 Rewards for Good Behavior and Punishments
for Poor Performance.........................................................258
7.9 How the Intelligent Computer Operates .........................................260
7.9.1 Sensing and Environmental Awareness .............................261
7.9.2 Decision-Making and Optimization...................................262
7.9.3 Case-Based Learning .........................................................262
7.9.4 Weight Values and Objective Functions ............................263
7.9.5 Distributed Learning ..........................................................263
7.10 Summary .........................................................................................263
Chapter 8: Cognitive Techniques: Position Awareness
John Polson and Bruce A. Fette ............................................... 269
8.1 Introduction.......................................................................................269
8.2 Radio Geolocation and Time Services ..............................................270
8.2.1 GPS ....................................................................................271
8.2.2 Coordinate System Transformations..................................275
8.2.3 GPS Geolocation Summary ...............................................275
8.3 Network Localization........................................................................276
8.3.1 Spatially Variant Network Service Availability .................276
8.3.2 Geolocation-Enabled Routing............................................278
8.3.3 Miscellaneous Functions....................................................278
8.4 Additional Geolocation Approaches .................................................278
8.4.1 Time-Based Approaches ....................................................279
8.4.2 AOA Approach ..................................................................286
x
Contents
8.5
8.6
8.7
8.8
8.9
8.4.3 RSS Approach....................................................................287
Network-Based Approaches..............................................................288
Boundary Decisions ..........................................................................288
8.6.1 Regulatory Region Selection .............................................288
8.6.2 Policy Servers and Regions................................................292
8.6.3 Other Uses of Boundary Decisions....................................293
Example of Cellular Telephone 911 Geolocation for First
Responders ........................................................................................293
Interfaces to Other Cognitive Technologies......................................294
8.8.1 Interface to Policy Engines ................................................294
8.8.2 Interface to Networking Functions ....................................295
8.8.3 Interface to Planning Engines ............................................295
8.8.4 Interface to User.................................................................295
Summary ...........................................................................................295
Chapter 9: Cognitive Techniques: Network
Awareness Jonathan M. Smith ........................................... 299
9.1
9.2
9.3
9.4
9.5
9.6
9.7
Introduction.......................................................................................299
Applications and their Requirements ................................................300
Network Solutions to Requirements .................................................302
Coping with the Complex Trade-Space ............................................304
Cognition to the Rescue ....................................................................306
The DARPA SAPIENT Program ......................................................308
Summary ...........................................................................................310
Chapter 10: Cognitive Services for the User
Joseph P.
Campbell, William M. Campbell, Scott M. Lewandowski
and Clifford J. Weinstein ....................................................... 313
10.1 Introduction.....................................................................................313
10.2 Speech and Language Processing ...................................................314
10.2.1 Speaker Recognition ........................................................314
10.2.2 Language Identification ...................................................323
10.2.3 Text-to-Speech Conversion ..............................................325
10.2.4 Speech-to-Text Conversion ..............................................325
10.2.5 Machine Translation ........................................................326
10.2.6 Background Noise Suppression .......................................327
10.2.7 Speech Coding .................................................................328
10.2.8 Speaker Stress Characterization.......................................329
xi
Contents
10.2.9 Noise Characterization.....................................................329
10.3 Concierge Services..........................................................................330
10.4 Summary .........................................................................................332
Chapter 11: Network Support: The Radio Environment
Map Youping Zhao, Bin Le and Jeffrey H. Reed ...................... 337
11.1
11.2
11.3
11.4
11.5
11.6
11.7
11.8
Introduction.....................................................................................337
Internal and External Network Support ..........................................338
Introduction to the REM .................................................................339
REM Infrastructure Support to Cognitive Radios...........................341
11.4.1 The Role of the REM in Cognitive Radio........................341
11.4.2 REM Design.....................................................................341
11.4.3 Enabling Techniques for Implementing REM .................343
Obtaining Awareness with the REM...............................................345
11.5.1 Awareness: Prerequisite for Cognitive Radio ..................345
11.5.2 Classification of Awareness .............................................347
11.5.3 Obtaining SA ...................................................................349
Network Support Scenarios and Applications ................................353
11.6.1 Infrastructure-Based Network and Centralized
Global REM .....................................................................354
11.6.2 Ad hoc Mesh Networks and Distributed
Local REMs .....................................................................355
Supporting Elements to the REM ...................................................357
Summary and Open Issues..............................................................360
Chapter 12: Cognitive Research: Knowledge Representation
and Learning Vincent J. Kovarik Jr. ..................................... 365
12.1 Introduction.....................................................................................365
12.2 Knowledge Representation and Reasoning.....................................369
12.2.1 Symbolic Representation .................................................371
12.2.2 Ontologies and Frame Systems........................................372
12.2.3 Behavioral Representation ...............................................374
12.2.4 Case-Based Reasoning.....................................................375
12.2.5 Rule-Based Systems.........................................................377
12.2.6 Temporal Knowledge.......................................................378
12.2.7 Knowledge Representation Summary..............................379
12.3 Machine Learning ...........................................................................380
12.3.1 Memorization...................................................................381
xii
Contents
12.3.2 Classifiers.........................................................................382
12.3.3 Bayesian Logic.................................................................383
12.3.4 Decision Trees..................................................................385
12.3.5 Reinforcement-Based Learning .......................................386
12.3.6 Temporal Difference ........................................................389
12.3.7 Neural Networks ..............................................................390
12.3.8 Genetic Algorithms..........................................................392
12.3.9 Simulation and Gaming ...................................................393
12.4 Implementation Considerations ......................................................394
12.4.1 Computational Requirements...........................................394
12.4.2 Brittleness and Edge Conditions......................................394
12.4.3 Predictable Behavior ........................................................395
12.5 Summary .........................................................................................396
Chapter 13: Roles of Ontologies in Cognitive Radios
Mieczyslaw M. Kokar, David Brady and Kenneth Baclawski ........... 401
13.1 Introduction to Ontology-Based Radio...........................................401
13.2 Knowledge-Intensive Characteristics of Cognitive Radio
401
13.2.1 Knowledge of Constraints and Requirements..................403
13.2.2 Information Collection and Fusion ..................................404
13.2.3 Situation Awareness and Advice......................................404
13.2.4 Self-awareness .................................................................405
13.2.5 Query by User, Self, or Other Radio................................405
13.2.6 Query Responsiveness and Command Execution ...........405
13.2.7 Negotiation for Resources................................................406
13.2.8 Dynamic Interoperability at Any Stack Layer .................406
13.3 Ontologies and Their Roles in Cognitive Radio .............................407
13.3.1 Introduction......................................................................407
13.3.2 Role of Ontology in Knowledge-Intensive
Applications .....................................................................413
13.4 A Layered Ontology and Reference Model ....................................414
13.4.1 Physical Layer Ontology..................................................414
13.4.2 Data Link Layer Ontology ...............................................416
13.5 Examples.........................................................................................421
13.5.1 Responding to Delays and Errors ....................................421
13.5.2 Adaptation of Training Sequence Length ........................423
13.5.3 Data Link Layer Protocol Consistency
and Selection....................................................................425
xiii
Contents
13.6 Open Research Issues .....................................................................427
13.6.1 Ontology Development and Consensus ...........................427
13.6.2 Ontology Mapping ...........................................................428
13.6.3 Learning ...........................................................................429
13.6.4 Efficiency of Reasoning ...................................................430
13.7 Summary .........................................................................................431
Chapter 14: Cognitive Radio Architecture
Joseph Mitola III ................................................................. 435
14.1 Introduction.....................................................................................435
14.2 CRA I: Functions, Components, and Design Rules........................436
14.2.1 AACR Functional Component Architecture....................436
14.2.2 Design Rules Include Functional Component
Interfaces..........................................................................441
14.2.3 Near-Term Implementations ............................................448
14.2.4 The Cognition Components .............................................450
14.2.5 Self-referential Components ............................................455
14.2.6 Flexible Functions of the Component Architecture.........458
14.3 CRA II: The Cognition Cycle .........................................................460
14.3.1 The Cognition Cycle ........................................................460
14.3.2 Observe (Sense and Perceive)..........................................461
14.3.3 Orient ...............................................................................462
14.3.4 Plan ..................................................................................463
14.3.5 Decide ..............................................................................464
14.3.6 Act....................................................................................464
14.3.7 Learning ...........................................................................464
14.3.8 Self-monitoring ................................................................465
14.4 CRA III: The Inference Hierarchy..................................................466
14.4.1 Atomic Stimuli.................................................................468
14.4.2 Primitive Sequences: Words and Dead Time ...................469
14.4.3 Basic Sequences...............................................................469
14.4.4 NL in the CRA Inference Hierarchy................................470
14.4.5 Observe–Orient Links for Scene Interpretation
472
14.4.6 Observe–Orient Links for Radio Skill Sets .....................473
14.4.7 General World Knowledge...............................................474
14.5 CRA IV: Architecture Maps ...........................................................476
14.5.1 CRA Topological Maps ...................................................477
14.5.2 CRA Identifies Self, Owner, and Home Network............478
xiv
Contents
14.6
14.7
14.8
14.9
14.5.3 CRA-Reinforced Hierarchical Sequences........................478
14.5.4 Behaviors in the CRA ......................................................479
14.5.5 From Maps to APIs..........................................................481
14.5.6 Industrial-Strength Inference Hierarchy ..........................481
CRA V: Building the CRA on SDR Architectures .........................483
14.6.1 Review of SWR and SDR Principles ...............................483
14.6.2 Radio Architecture ...........................................................486
14.6.3 The SCA...........................................................................487
14.6.4 Functions-Transforms Model of Radio............................490
14.6.5 Architecture Migration: From SDR to AACR.................491
14.6.6 Cognitive Electronics.......................................................491
14.6.7 When Should a Radio Transition toward Cognition? ......492
14.6.8 Radio Evolution toward the CRA ....................................494
Cognition Architecture Research Topics ........................................494
Industrial-Strength AACR Design Rules........................................495
Summary and Future Directions .....................................................497
Chapter 15: Cognitive Radio Performance Analysis James O. Neel,
Jeffrey H. Reed and Allen B. MacKenzie .................................... 501
15.1 Introduction.....................................................................................501
15.2 The Analysis Problem.....................................................................505
15.2.1 Mathematical Preliminaries .............................................505
15.2.2 A Formal Model of a Cognitive Radio Network .............506
15.2.3 Analysis Objectives..........................................................509
15.3 Traditional Engineering Analysis Techniques ................................513
15.3.1 A Dynamical Systems Approach .....................................513
15.3.2 Contraction Mappings and the General Convergence
Theorem ...........................................................................518
15.3.3 Markov Models ................................................................524
15.4 Applying Game Theory to the Analysis Problem...........................529
15.4.1 Basic Elements of Game Theory .....................................530
15.4.2 Mapping the Basic Elements of a Game
to the Cognition Cycle .....................................................533
15.4.3 Basic Game Models .........................................................534
15.4.4 Basic Game Theory Analysis Techniques .......................538
15.5 Relevant Game Models ...................................................................544
15.5.1 Potential Games ...............................................................544
15.5.2 Supermodular Games.......................................................554
xv
Contents
15.6 Case Studies ....................................................................................563
15.6.1 Distributed Power Control ...............................................563
15.6.2 Dynamic Frequency Selection .........................................568
15.6.3 Adaptive Interference Avoidance.....................................569
15.7 Summary and Conclusions .............................................................572
15.8 Questions.........................................................................................575
Chapter 16: The Really Hard Problems
Bruce A. Fette .............. 581
16.1 Introduction.....................................................................................581
16.2 Review of the Book.........................................................................581
16.3 Services Offered to Wireless Networks through Infrastructure ......587
16.3.1 Stand-Alone Radios with Cognition ................................588
16.3.2 Cellular Infrastructure Support to Cognition ...................589
16.3.3 Data Radios......................................................................590
16.3.4 Cognitive Services Offered through Infrastructure..........591
16.3.5 The Remaining Hard Problems........................................593
Glossary ................................................................................. 595
Index...................................................................................... 609
xvi
List of Contributors
William M. Campbell
Information Systems Technology
Group
MIT Lincoln Laboratory
244 Wood Street, C-243
Lexington, MA, 02420-9185
Email: [email protected]
Kenneth Baclawski
Computer and Information Science
Northeastern University
360 Huntington Avenue
Boston, MA, 02115
Charles W. Bostian
Wireless @ Virginia Tech
Bradley Department of Electrical
and Computer Engineering
Virginia Tech
Mail Code 0111
Blacksburg, VA, 24060-0111
Email: [email protected]
Bruce A. Fette
Chief Scientist
Communication Networks Division
General Dynamics C4 Systems
8220 E Roosevelt
Scottsdale, AZ, 85257
Email: [email protected]
David Brady
ECE Dept
Northeastern University
360 Huntington Avenue
Boston MA, 02115
Email: [email protected]
Mieczyslaw M. Kokar
Department of Electrical and
Computer Engineering
Northeastern University
360 Huntington Avenue
Boston, MA, 02115
Email: [email protected]
Joseph P. Campbell
Senior MTS
Information Systems Technology
Group
MIT Lincoln Laboratory
244 Wood Street, C-290A
Lexington, MA, 02420-9185
Email: [email protected]
Paul Kolodzy
Kolodzy Consulting
P.O. Box 1443
Centerville, VA, 20120
Email: [email protected]
xvii
List of Contributors
Vincent J. Kovarik Jr.
Harris Corporation
Mail Stop W2-11F
P.O. Box 37
Melbourne FL, 32902-0037
Email: [email protected]
Bin Le
Center for Wireless
Telecommunications
Wireless @ Virginia Tech
Bradley Department of Electrical
and Computer Engineering
Virginia Tech, Mail Code 0111
Blacksburg, VA, 24061-0111
Email: [email protected]
Scott M. Lewandowski
Information Systems Technology
Group
MIT Lincoln Laboratory
244 Wood Street, C-256
Lexington, MA, 02420-9185
Email: [email protected]
Preston Marshall
Defense Advanced Research Projects
Agency
Email: [email protected]
Allen B. MacKenzie
Bradley Department of Electrical
and Computer Engineering
Virginia Polytechnic Institute and
State University, Mail Code 0111
Blacksburg, VA, 24061-0111
Email: [email protected]
Joseph Mitola III
Consulting Scientist
Tampa, FL, 33604
Email: [email protected]
James O. Neel
Mobile and Portable Radio Research
Group
Wireless @ Virginia Tech
Bradley Department of Electrical
and Computer Engineering
Virginia Tech
432 Durham Hall, MS 0350
Blacksburg, VA, 24061
Email: [email protected]
John Polson
Principal Engineer
Bell Helicopter, Textron Inc.
P.O. Box 482
Fort Worth, TX, 76101
Email: jtpolson@bellhelicopter.
textron.com
Jeffrey H. Reed
Mobile and Portable Radio Research
Group
Wireless @ Virginia Tech
Bradley Department of Electrical
and Computer Engineering
Virginia Tech
432 Durham Hall, MS 350
Blacksburg, VA, 24061
Email: [email protected]
xviii
List of Contributors
Clifford J. Weinstein
Group Leader
Information Systems Technology
Group
MIT Lincoln Laboratory
244 Wood Street, C-290A
Lexington, MA, 02420-9185
Email: [email protected]
Pablo Robert
Mobile and Protable Radio Research
Group (MPRG)
Bradley Department of Electrical
and Computer Engineering
Virginia Tech
Blacksburg, VA, 24060-0111
Email: [email protected]
Robert J. Wellington
University of Minnesota
9740 Russel Circle S.
Bloomington MN, 55431
Email: [email protected]
Thomas W. Rondeau
Bradley Department of Electrical
and Computer Engineering
Virginia Tech
Mail Code 0111
Blacksburg, VA, 24060-0111
Email: [email protected]
Youping Zhao
Mobile and Portable Radio Research
Group
Wireless @ Virginia Tech
432 Durham Hall, MS 350
Blacksburg, VA, 24061
Email:[email protected]
Jonathan M. Smith
Defense Advanced Research Projects
Agency
Email: [email protected]
xix
This page intentionally left blank
Foreword
This introduction takes a visionary look at ideal cognitive radios (CRs) that integrate advanced software-defined radios (SDR) with CR techniques to arrive at
radios that learn to help their user using computer vision, high-performance
speech understanding, global positioning system (GPS) navigation, sophisticated
adaptive networking, adaptive physical layer radio waveforms, and a wide range
of machine learning processes.
CRs Know Radio Like TellMe® Knows 800 Numbers
When you dial 1-800-555-1212, a speech synthesis algorithm says “Toll Free
Directory Assistance powered by TellMe® Networks (www.tellme.com, Mountain
View, CA, 2005). Please say the name of the listing you want.” If you mumble it
says, “OK, United Airlines. If that is not what you wanted press 9, otherwise wait
while I look up the number.” Reportedly, some 99 percent of the time TellMe gets
it right, replacing the equivalent of thousands of directory assistance operators of
yore. TellMe, a speech-understanding system, achieves a high degree of success
by its focus on just one task: finding a toll-free telephone number. Narrow task
focus is one key to algorithm successes.
The cognitive radio architecture (CRA) is the building block from which to
build cognitive wireless networks (CWNs), the wireless mobile offspring of
TellMe. CRs and networks are emerging as practical, real-time, highly focused
applications of computational intelligence technology. CRs differ from the more
general artificial intelligence (AI) based services like intelligent agents, computer
speech, and computer vision in degree of focus. Like TellMe, CRs focus on
very narrow tasks. For CRs, the task is to adapt radio-enabled information services to the specific needs of a specific user. TellMe, a network service, requires
Note: Adapted from J. Mitola III, Aware, Adaptive and Cognitive Radio: The Engineering
Foundations of Radio XML, Wiley, 2006.
xxi
Foreword
substantial network computing resources to serve thousands of users at once.
CWNs, on the other hand, may start with a radio in your purse or on your belt, a
cell phone on steroids, focused on the narrow task of creating from the myriad
available wireless information networks and resources just what is needed by just
one user: you. Each CR fanatically serves the needs and protects the personal
information of just one owner via the CRA using its audio and visual sensory perception and automated machine learning (AML).
TellMe is here and now, while CRs are emerging in global wireless research
centers and industry forums like the SDR Forum and Wireless World Research
Forum (WWRF). This book introduces the technologies to bootstrap CR systems,
introducing technical challenges and approaches, emphasizing CR as a technology
enabler for rapidly emerging commercial CWN services.
CRs See What You See, Discovering Radio Frequency Uses,
Needs, and Preferences
Although the common cell phone may have a camera, it lacks vision algorithms,
so it does not know what it is seeing. It can send a video clip, but it has no
perception of the visual scene in the clip. If it had vision-processing algorithms,
it could perceive and understand the visual scene. It could tell whether it were at
home, in the car, at work, shopping, or driving up the driveway on the way
home. If vision algorithms show it that you are entering your driveway in your
car, a CR could learn to open the garage door for you wirelessly. Thus, you
would not need to fish for the garage door opener, yet another wireless gadget.
In fact, you do not need a garage door opener anymore, once CRs enter the
market. To open the car door, you will not need a key fob either. As you approach
your car, your personal CR perceives the common scene and, as trained,
synthesizes the fob radio frequency (RF) transmission and opens the car door
for you.
CRs do not attempt everything. They learn about your radio use patterns
because they know a lot about radio, generic users, and legitimate uses of radio.
CRs have the a priori knowledge needed to detect opportunities to assist you with
your use of the radio spectrum accurately, delivering that assistance with minimum intrusion.
Products realizing the visual perception of this vignette are demonstrated
on laptop computers today. Reinforcement learning (RL) and case-based reasoning (CBR) are mature AML technologies with radio network applications now
being demonstrated in academic and industrial research settings as technology
xxii
Foreword
pathfinders for CR1 and CWN.2 Two or three Moore’s law cycles or 3 to 5 years
from now, these vision and learning algorithms will fit in your cell phone. In the
interim, CWNs will begin to offer such services, presenting consumers with new
tradeoffs between privacy and ultra-personalized convenience.
CRs Hear What You Hear, Augmenting Your Personal Skills
The cell phone on your waist is deaf. Although your cell phone has a microphone,
it lacks embedded speech-understanding technology, so it does not perceive what
it hears. It can let you talk to your daughter, but it has no perception of your
daughter, nor of the content of your conversation. If it had speech-understanding
technology, it could perceive your speech dialog. It could detect that you and your
daughter are talking about common subjects like homework or your favorite song.
With CR, speech algorithms would detect your daughter saying that your favorite
song is now playing on WDUV. As an SDR, not just a cell phone, your CR then
could tune to FM 105.5 so that you can hear “The Rose.”
With your CR, you no longer need a transistor radio. Your CR eliminates from
your pocket, purse or backpack yet another RF gadget. In fact, you may not need
iPOD®, Game Boy® and similar products as high-end CRs enter the market (or
iPODs or Game Boys with CR may become the single pocket pal instead: you never
know how market demand will shape products toward the “killer app,” do you?).
Your CR could learn your radio listening and information use patterns, accessing
songs, downloading games, snipping broadcast news, sports, and stock quotes as
you like as the CR re-programs its internal SDR to better serve your needs and
preferences. Combining vision and speech perception, as you approach your car your
CR perceives this common scene and, as you had the morning before, tunes your car
radio to WTOP for your favorite “Traffic and weather together on the eights.”
1
Mitola’s reference for CR pathfinders.
Semantic Web: Researchers formulate CRs as sufficiently speech-capable to answer questions
about <Self/> and the <Self/> use of <Radio/> in support of its <Owner/>. When an ordinary concept like “owner” has been translated into a comprehensive ontological structure of Computational
primitives, for example, via Semantic Web technology, the concept becomes a computational
primitive for autonomous reasoning and information exchange. Radio XML, an emerging CR
derivative of the eXtensible Markup Language, XML, offers to standardize such radio-scene perception primitives. They are highlighted in this brief treatment by <Angle-brackets/>. All CR have
a <Self/>, a <Name/>, and an <Owner/>. The <Self/> has capabilities like <GSM/> and <SDR/>,
a self-referential computing architecture, which is guaranteed to crash unless its computing
ability is limited to real-time response tasks; this is appropriate for CR but may be too limiting
for general-purpose computing.
2
xxiii
Foreword
For AML, CRs need to save speech, RF, and visual cues, all of which may be
recalled by the user, expanding the user’s ability to remember details of conversations and snapshots of scenes, augmenting the skills of the <owner/>.3 Because of
the brittleness of speech and vision technologies, CRs try to “remember everything” like a continuously running camcorder. Since CRs detect content, such as
speakers’ names, and keywords like “radio” and “song,” they can retrieve some
content asked for by the user, expanding the user’s memory in a sense. CRs thus
could enhance the personal skills of their users, such as memory for detail.
CRs Learn to Differentiate Speakers to Reduce Confusion
To further limit combinatorial explosion in speech, CR may form speaker models,
statistical summaries of the speech patterns of speakers, particularly of the
<Owner/>. Speaker modeling is particularly reliable when the <Owner/> uses the
CR as a cell phone to place a phone call. Contemporary speaker recognition algorithms differentiate male from female speakers with high accuracy. With a few
different speakers to be recognized (i.e., fewer than 10 in a family) and with reliable side information like the speaker’s telephone number, today’s state-of-the-art
algorithms recognize individual speakers with better than 95 percent accuracy.
Over time, each CR learns the speech patterns of its <Owner/> in order to
learn from the <Owner/> and not be confused by other speakers. CR thus leverages experience incrementally to achieve increasingly sophisticated dialog. Today,
a 3 GHz laptop supports this level of speech understanding and dialog synthesis in
real time, making it likely to be available in a cell phone in 3 to 5 years.
The CR must both know a lot about radio and learn a lot about you, the
<Owner/>, recording and analyzing personal information and thus placing a premium on trustworthy privacy technologies. Increased autonomous customization
of wireless service include secondary use of broadcast spectrum. Therefore, the
CRA incorporates speech recognition to enable learning without requiring overwhelming amounts of training, allowing it to become sufficiently helpful without
being a nuisance.
More Flexible Secondary Use of Radio Spectrum
In 2004, the US Federal Communications Commission (FCC) issued a Report and
Order that radio spectrum allocated to TV, but unused in a particular broadcast
3
Ibid.
xxiv
Foreword
market, such as a rural area, could be used by CR as secondary users under Part
15 rules for low-power devices—for example, to create ad hoc networks. SDR
Forum member companies have demonstrated CR products with these elementary
spectrum perception and use capabilities. Wireless products—both military and
commercial—are realizing that the FCC vignettes already exist.
Complete visual and speech perception capabilities are not many years distant.
Productization is underway. Thus, many chapters of Bruce’s outstanding book
emphasize CR spectrum agility, suggesting pathways toward enhanced perception
technologies, with new long-term growth paths for the wireless industry. This
book’s contributors hope that it will help you understand and create new opportunities for CR technologies.
Dr. Joseph Mitola III
Tampa, Florida
xxv
This page intentionally left blank
CHAPTER 1
History and Background of
Cognitive Radio Technology
Bruce A. Fette
Communications Networks Division
General Dynamics C4 Systems
Scottsdale, AZ, USA
1.1 The Vision of Cognitive Radio
Just imagine if your cellular telephone, personal digital assistant (PDA), laptop,
automobile, and TV were as smart as “Radar” O’Reilly from the popular TV series
M*A*S*H.1 They would know your daily routine as well as you do. They would
have things ready for you as soon as you ask, almost in anticipation of your need.
They would help you find people, things, and opportunities; translate languages; and
complete tasks on time. Similarly, if a radio were smart, it could learn services available in locally accessible wireless computer networks, and could interact with those
networks in their preferred protocols, so you would have no confusion in finding
the right wireless network for a video download or a printout. Additionally, it
could use the frequencies and choose waveforms that minimize and avoid interference with existing radio communication systems. It might be like having a friend
in everything that’s important to your daily life, or like you were a movie director
with hundreds of specialists running around to help you with each task, or like
you were an executive with hundred assistants to find documents, summarize them
into reports, and then synopsize the reports into an integrated picture. A cognitive
radio is the convergence of the many pagers, PDAs, cell phones, and many other
1
“Radar” O’Reilly is a character in the popular TV series M*A*S*H, which ran from 1972 to
1983. He always knew what the colonel needed before the colonel knew he needed it.
1
Chapter 1
single-purpose gadgets we use today. They will come together over the next
decade to surprise us with services previously available to only a small select
group of people, all made easier by wireless connectivity and the Internet.
1.2 History and Background Leading to Cognitive Radio
The sophistication possible in a software-defined radio (SDR) has now reached the
level where each radio can conceivably perform beneficial tasks that help the user,
help the network, and help minimize spectral congestion. Radios are already
demonstrating one or more of these capabilities in limited ways. A simple example
is the adaptive digital European cordless telephone (DECT) wireless phone, which
finds and uses a frequency within its allowed plan with the least noise and interference on that channel and time slot. Of these capabilities, conservation of spectrum
is already a national priority in international regulatory planning. This book leads
the reader through the technologies and regulatory considerations to support three
major applications that raise an SDR’s capabilities and make it a cognitive radio:
1. Spectrum management and optimizations.
2. Interface with a wide variety of networks and optimization of network resources.
3. Interface with a human and providing electromagnetic resources to aid the
human in his or her activities.
Many technologies have come together to result in the spectrum efficiency and
cognitive radio technologies that are described in this book. This chapter gives the
reader the background context of the remaining chapters of this book. These technologies represent a wide swath of contributions upon which cognitive technologies may be considered as an application on top of a basic SDR platform.
To truly recognize how many technologies have come together to drive cognitive radio techniques, we begin with a few of the major contributions that have led
up to today’s cognitive radio developments. The development of digital signal
processing (DSP) techniques arose due to the efforts of such leaders as Alan
Oppenheim [1], Lawrence Rabiner [2, 3], Ronald Schaefer [3], Ben Gold, Thomas
Parks [4], James McClellen [4], James Flanagan [5], fred harris [6], and James
Kaiser. These pioneers2 recognized the potential for digital filtering and DSP,
and prepared the seminal textbooks, innovative papers, and breakthrough signal
2
This list of contributors is only a partial representative listing of the pioneers with whom the
author is personally familiar, and not an exhaustive one.
2
History and Background of Cognitive Radio Technology
processing techniques to teach an entire industry how to convert analog signal
processes to digital processes. They guided the industry in implementing new
processes that were entirely impractical in analog signal processing.
Somewhat independently, Cleve Moler, Jack Little, John Markel, Augustine
Gray, and others began to develop software tools that would eventually converge
with the DSP industry to enable efficient representation of the DSP techniques,
and would provide rapid and efficient modeling of these complex algorithms [7, 8].
Meanwhile, the semiconductor industry, continuing to follow Moore’s law [9],
evolved to the point where the computational performance required to implement
digital signal processes used in radio modulation and demodulation were not only
practical, but resulted in improved radio communication performance, reliability,
flexibility, and increased value to the customer. This meant that analog functions
implemented with large discrete components were replaced with digital functions
implemented in silicon, and consequently were more producible, less expensive,
more reliable, smaller, and of lower power [10].
During this same period, researchers all over the globe explored various techniques to achieve machine learning and related methods for improved machine
behavior. Among these were analog threshold logic, which lead to fuzzy logic and
neural networks, a field founded by Frank Rosenblatt [11]. Similarly, languages to
express knowledge and to understand knowledge databases evolved from list
processing (LISP) and Smalltalk and from massive databases with associated
probability statistics. Under funding from the Defense Advanced Research
Projects Agency (DARPA), many researchers worked diligently on understanding
natural language and understanding spoken speech. Among the most successful
speech-understanding systems were those developed by Janet and Jim Baker (who
subsequently founded Dragon Systems) [12], and Kai Fu Lee et al. [13]. Both of
these systems were developed under the mentoring of Raj Reddy at Carnegie
Mellon. Today, we see Internet search engines reflecting the advanced state of
artificial intelligence (AI).
In networking, DARPA and industrial developers at Xerox, BBN Technologies,
IBM, ATT, and Cisco each developed computer-networking techniques, which
evolved into the standard Ethernet and Internet we all benefit from today. The
Internet Engineering Task Force (IETF), and many wireless-networking researchers
continue to evolve networking technologies with a specific focus on making radio
networking as ubiquitous as our wired Internet. These researchers are exploring
wireless networks that range from access directly via a radio access point to more
advanced techniques in which intermediate radio nodes serve as repeaters to forward data packets toward their eventual destination in an ad hoc network topology.
3
Chapter 1
All of these threads come together as we arrive today at the cognitive radio era
(see Figure 1.1). Cognitive radios are nearly always applications that sit on top of
an SDR, which in turn is implemented largely from digital signal processors and
general-purpose processors (GPPs) built in silicon. In many cases, the spectral
efficiency and other intelligent support to the user arises by sophisticated networking of many radios to achieve the end behavior, which provides added capability
and other benefits to the user.
Semiconductor Processor, DSP,
A/D, and D/A Architectures
Wireless
Networking
AI Languages
and Knowledge
Databases
Cognitive Radio
Network Infrastructure
Math and Signal
Processing Tool
Development
Source Coding of
Speech, Imagery,
Video, and Data
Cognitive Radio
Protocols and
Etiquettes
Basic
SDR
DSP
Technologies
The
Ultimate
Cognitive
Radio
Regulatory
Support
1970s 1980s 1990s 2000s
Standardized
Cognitive Radio
Architecture
Cognitive Radio
Business Model
2006
Figure 1.1: Technology timeline. Synergy among many technologies converge to enable the
SDR. In turn, the SDR becomes the platform of choice for the cognitive radio.
1.3 A Brief History of SDR
An SDR is a radio in which the properties of carrier frequency, signal bandwidth,
modulation, and network access are defined by software. Today’s modern SDR
also implements any necessary cryptography; forward error correction (FEC) coding; and source coding of voice, video, or data in software as well. As shown in
the timeline of Figure 1.2, the roots of SDR design go back to 1987, when Air
Force Rome Labs (AFRL) funded the development of a programmable modem as
an evolutionary step beyond the architecture of the integrated communications,
navigation, and identification architecture (ICNIA). ICNIA was a federated design
of multiple radios, that is, a collection of several single-purpose radios in one box.
4
History and Background of Cognitive Radio Technology
DMR
ICNIA (Rx, Tx)
1970
1980
SPEAKeasy-I
SE-I
Demo
1990
Cluster
HMS
SPEAKeasy-II
SE-II
Demo
1995
1997
MMITS/
SDR Forum
2004
JTRS
JPO
Stood up
Figure 1.2: SDR timeline. Images of ICNIA, SPEAKeasy I (SE-I), SPEAKeasy II (SE-II), and
Digital Modular Ratio (DMR) on their contract award timelines and corresponding
demonstrations. These radios are the early evolutionary steps that lead to today’s SDR.
Today’s SDR, in contrast, is a general-purpose device in which the same radio
tuner and processors are used to implement many waveforms at many frequencies.
The advantage of this approach is that the equipment is more versatile and costeffective. Additionally, it can be upgraded with new software for new waveforms
and new applications after sale, delivery, and installation. Following the programmable modem, AFRL and DARPA joined forces to fund the SPEAKeasy I and
SPEAKeasy II programs.
SPEAKeasy I was a six-foot-tall rack of equipment (not easily portable), but it
did demonstrate that a completely software-programmable radio could be built,
and included a software-programmable cryptography chip called Cypress, with software cryptography developed by Motorola (subsequently purchased by General
Dynamics). SPEAKeasy II was a complete radio packaged in a practical radio size
(the size of a stack of two pizza boxes), and was the first SDR to include programmable voice coder (vocoder), and sufficient analog and DSP resources to handle many
different kinds of waveforms. It was subsequently tested in field conditions at Ft.
Irwin, California, where its ability to handle many waveforms underlined its
extreme usefulness, and its construction from standardized commercial off-the-shelf
(COTS) components was a very important asset in defense equipment. SPEAKeasy
II subsequently evolved into the US Navy’s digital modular radio (DMR), becoming
a four-channel full duplex SDR, with many waveforms and many modes, able to
5
Chapter 1
WF (d)
WF (c)
WF (b)
WF (a)
WF User Application (Vocoder, Browser, etc.)
WF Network Layer
Waveform(n)
WF MAC Layer
Standardized OS Interface (POSIX Compliance Shim)
Operating System
APIs
Standardized
Interfaces
Security Service Drivers
Multiprocessor
Intercommunication
Infrastructure (CORBA
or Equivalent)
Radio Services Drivers
Software Communication
Architecture Core Framework
Wireless Network Services Drivers
WF PHY Layer
Board Support: Basic HW Drivers, Boot, BIST
Hardware Components and Processors
Multiple Processor Resources
Figure 1.3: Basic software architecture of a modern SDR.3 Standardized APIs are defined
for the major interfaces to assure software portability across many very different hardware
platform implementations. The software has the ability to allocate computational resources
to specific waveforms. It is normal for an SDR to support many waveforms in order to
interface to many networks, and thus to have a library of waveforms and protocols.
be remotely controlled over an Ethernet interface using Simple Network Management Protocol (SNMP).
These SPEAKeasy II and DMR products evolved not only to define these
radio waveform features in software, but also to develop an appropriate software
architecture to enable porting the software to an arbitrary hardware platform, and
thus to achieve hardware independence of the waveform software specification.
This critical step allows the hardware to separately evolve its architecture independently from the software, and thus frees the hardware to continue to evolve
and improve after delivery of the initial product.
The basic hardware architecture of a modern SDR (Figure 1.3) provides sufficient resources to define the carrier frequency, bandwidth, modulation, any
3
BIST: built-in self-test; CORBA: Common Object Request Broker Architecture; HW: hardware;
MAC: medium access control; OS: operating system; PHY: physical (layer); POSIX: Portable
Operating System Interface; WF: waveform.
6
History and Background of Cognitive Radio Technology
Duplexer and
Antenna Manager/Tuner
RF Front-End
Tunable Filters
and LNA
Digital Back-End
IF/AGC
Mixer
A/D
Mixer
GPPs
User Interface
Peripherals
FPGAs
Local Oscillators
Tunable Filters and
Power Amplifier
DSPs
Carrier
Synthesizer
D/A
Specialized
Co-processors
Power
Manager
Figure 1.4: Basic hardware architecture of a modern SDR.4 The hardware provides sufficient
resources to define the carrier frequency, bandwidth, modulation, any necessary cryptography, and source coding in software. The hardware resources may include mixtures of GPPs,
DSPs, FPGAs, and other computational resources, sufficient to include a wide range of
modulation types.
necessary cryptography, and source coding in software. The hardware resources
may include mixtures of GPPs, DSPs, field-programmable gate arrays (FPGAs),
and other computational resources, sufficient to include a wide range of modulation types (see Section 1.4.1). In the basic software architecture of a modern SDR
(Figure 1.4), the application programming interfaces (APIs) are defined for the
major interfaces to assure software portability across many very different hardware platform implementations, as well as to assure that the basic software infrastructure supports a wide diversity of waveform applications without having to be
rewritten for each waveform or application. The software has the ability to allocate computational resources to specific waveforms (see Section 1.4.2). It is normal for an SDR to support many waveforms in order to interface to many
networks, and thus to have a library of waveforms and protocols.
The SDR Forum was founded in 1996 by Wayne Bonser of AFRL to develop
industry standards for SDR hardware and software that could assure that the software not only ports across various hardware platforms, but also defines standardized interfaces to facilitate porting software across multiple hardware vendors and
to facilitate integration of software components from multiple vendors. The SDR
Forum is now a major influence in the SDR industry, dealing not only with standardization of software interfaces but many other important enabling technology
4
A/D: analog to digital; AGC: automatic gain control; D/A: digital to analog; IF: intermediate
frequency; LNA: low-noise amplifier; RF: radio frequency.
7
Chapter 1
issues in the industry from tools, to chips, to applications, to cognitive radio and
spectrum efficiency. The SDR Forum currently has a Cognitive Radio Working
Group, which is preparing papers to advance both spectrum efficiency and cognitive radio applications. In addition, special interest groups within the Forum have
interests in these topics.
The SDR Forum Working Group is treating cognitive radio and spectrum efficiency as applications that can be added to an SDR. This means that we can begin
to assume an SDR as the basic platform upon which to build most new cognitive
radio applications.
1.4 Basic SDR
In this section, we endeavor to provide the reader with background material to
provide a basis for understanding subsequent chapters.
1.4.1 The Hardware Architecture of an SDR
The basic SDR must include the radio front-end, the modem, the cryptographic
security function, and the application function. In addition, some radios will also
include support for network devices connected to either the plain text side or the
modem side of the radio, allowing the radio to provide network services and to be
remotely controlled over the local Ethernet.
Some radios will also provide for control of external radio frequency (RF)
analog functions such as antenna management, coax switches, power amplifiers,
or special-purpose filters. The hardware and software architectures should allow
RF external features to be added if or when required for a particular installation or
customer requirement.
The RF front-end (RFFE) consists of the following functions to support the
receive mode: antenna-matching unit, low-noise amplifier, filters, local oscillators,
and analog-to-digital (A/D) converters (ADCs) to capture the desired signal and
suppress undesired signals to a practical extent. This maximizes the dynamic
range of the ADC available to capture the desired signal.
To support the transmit mode, the RFFE will include digital-to-analog (D/A)
converters (DACs), local oscillators, filters, power amplifiers, and antenna-matching
circuits. In transmit mode, the important property of these circuits is to synthesize
the RF signal without introducing noise and spurious emissions at any other frequencies that might interfere with other users in the spectrum.
8
History and Background of Cognitive Radio Technology
AGC
A/D DC Offset
I/Q Balance
Coarse Filter
Frequency Offset
De-spread
Fine Filter
Fine Baud Timing
Interference
Suppressor
Channel
Equalizer
Soft Decision
Demodulator
Tracking Loops
Parameter
Estimators
Inner FEC
De-multiplexer
Outer FEC
Networking Control
Message Analysis
Bits to Application
Layer
Figure 1.5: Traditional digital receiver signal processing block diagram.5
The modem processes the received signal or synthesizes the transmitted signal, or both for a full duplex radio. In the receive process (Figure 1.5), the modem
will shift the carrier frequency of the desired signal to a specific frequency nearly
equivalent to heterodyne shifting the carrier frequency to direct current (DC), as
perceived by the digital signal processor, to allow it to be digitally filtered. The
digital filter provides a high level of suppression of interfering signals not within
the bandwidth of the desired signal. The modem then time-aligns and de-spreads
the signal as required, and refilters the signal to the information bandwidth. Next the
modem time-aligns the signal to the symbol or baud time so that it can optimally
align the demodulated signal with expected models of the demodulated signal.
The modem may include an equalizer to correct for channel multipath artifacts,
and for filtering and delay distortions. It may also optionally include rake filtering
to optimally cohere multipath components for demodulation. The modem will
compare the received symbols with the possible received symbols and make a best
possible estimate of which symbols were transmitted. Of course, if there is a weak
signal or strong interference, some symbols may be received in error. If the waveform includes FEC coding, the modem will decode the received sequence of
encoded symbols by using the structured redundancy introduced in the coding
process to detect and correct the encoded symbols that were received in error.
5
I/Q, meaning “in phase and quadrature,” is the real part and the imaginary part of the complexvalued signal after being sampled by the ADC(s) in the receiver, or as synthesized by the modem
and presented to the DAC in the transmitter.
9
Chapter 1
The process the modem performs for transmit (Figure 1.6) is the inverse of
that for receive. The modem takes bits of information to be transmitted, groups the
information into packets, adds a structured redundancy to provide for error correction at the receiver, groups bits to be formed into symbols, selects a wave shape to
represent each symbol, synthesizes each wave shape, and filters each wave shape
to keep it within its desired bandwidth. It may spread the signal to a much wider
bandwidth by multiplying the symbol by a wideband waveform which is also generated by similar methods. The final waveform is filtered to match the desired
transmit signal bandwidth. If the waveform includes a time-slotted structure such
as time division multiple access (TDMA) waveforms, the radio will wait for the
appropriate time while placing samples that represent the waveform into an output
first in, first out (FIFO) buffer ready to be applied to the DAC. The modem must
also control the power amplifier and the local oscillators to produce the desired
carrier frequency, and must control the antenna-matching unit to minimize voltage
standing wave radio (VSWR). The modem may also control the external RF elements, including transmit versus receive mode, carrier frequency, and smart antenna
control. Considerable detail on the architecture of SDR is given by Reed [14].
Bits from
Application
Layer
Networking
Control
Message
Analysis
Modulator
Multiplexer
and Optional
Cryptography
Spectral
Shaping
Outer FEC
Optional
Spreading,
Optional
Shaping
Inner FEC
Predistortion
PA Compensation
Bit to Symbol
Mapping
Queuing
for Media
Access
Control
Transmit
I/Q
Waveform
to D/A
Figure 1.6: Traditional transmit signal processing block diagram.
The cryptographic security function must encrypt any information to be transmitted. Because the encryption processes are unique to each application, these
cannot be generalized. The Digital Encryption Standard (DES) and the Advanced
Encryption Standard (AES) from the US National Institute of Standards and
Technology (NIST) provide example cryptographic processes [15, 16]. In addition
to providing the user with privacy for voice communication, cryptography also
plays a major role in assuring that the billing is to an authenticated user terminal.
10
History and Background of Cognitive Radio Technology
In the future, it will also be used to authenticate transactions of delivering software and purchasing services. In future cognitive radios, the policy functions that
define the radios’ allowed behaviors will also be cryptographically sealed to prevent tampering with regulatory policy as well as network operator policy.
The application processor will typically implement a vocoder, a video coder,
and/or a data coder, as well as selected web browser functions. In each case, the
objective is to use knowledge of the properties of the digitized representation of
the information to compress the data rate to an acceptable level for transmission.
Voice, video, and data coding typically utilize knowledge of the redundancy in the
source signal (speech or image) to compress the data rate. Compression factors
typically in excess of 10:1 are achieved in voice coding, and up to 100:1 in video
coding. Data coding has a variety of redundancies within the message, or between
the message and common messages sent in that radio system. Data compression
ranges from 10 to 50 percent, depending on how much redundancy can be identified in the original information data stream.
Typically, speech and video applications run on a DSP processor. Text and
web browsing typically run on a GPP. As speech recognition technology continues to improve its accuracy, we can expect that the keyboard and display will be
augmented by speech input and output functionality. On cognitive radios with
adequate processors, it may be possible to run speech recognition and synthesis
on the cognitive radio, but early units may find it preferable to vocode the voice,
transmit the voice to the base station, and have recognition and synthesis performed at an infrastructure component. This will keep the complexity of the
portable units smaller.
1.4.2 Computational Processing Resources in an SDR
The design of an SDR must anticipate the computational resources needed to
implement its most complex application. The computational resources may
consist of GPPs, DSPs, FPGAs, and occasionally will include other chips that
extend the computational capacity. Generally, the SDR vendor will avoid inclusion of dedicated-purpose non-programmable chips because the flexibility to
support waveforms and applications is limited, if not rigidly fixed, by nonprogrammable chips.
Currently, an example GPP selected by many SDR developers is the PowerPC.
The PowerPC is available from several vendors. This class of processor is readily
programmed in standard C or C language, supports a very wide variety of
addressing modes, floating point and integer computation, and a large memory
11
Chapter 1
space, usually including multiple levels of on-chip and off-chip cache memory.
These processors currently perform more than 1 billion mathematical operations
per second (mops).6 GPPs in this class usually pipeline the arithmetic functions
and decision logic functions several levels deep in order to achieve these speeds.
They also frequently execute many instructions in parallel, typically performing
the effective address computations in parallel with arithmetic computation, logic
evaluations, and branch decisions.
Most important to the waveform modulation and demodulation processes is
the speed at which these processors can perform real or complex multiply accumulates. The waveform signal processing represents more than 90 percent of the
total computational load in most waveforms, although the protocols to participate
in the networks frequently represent 90 percent of the lines of code. Therefore,
it is of great importance to the hardware SDR design that the SDR architecture
include DSP-type hardware multiply accumulate functions, so that the signal
processes can be performed at high speed, and GPP-type processors for the
protocol stack processing.
DSPs are somewhat different than GPPs. The DSP internal architecture is
optimized to be able to perform multiply accumulates very fast. This means they
have one or more multipliers and one or more accumulators in hardware. Usually
the implication of this specialization is that the device has a somewhat unusual
memory architecture, usually partitioned so that it can fetch two operands simultaneously and also be able to fetch the next software instruction in parallel with the
operand fetches. Currently, DSPs are available that can perform fractional mathematics (integer) multiply accumulate instructions at rates of 1 GHz, and floating
point multiply accumulates at 600 MHz. DSPs are also available with many parallel multiply accumulate engines, reporting rates of more than 8 Gmops. The other
major feature of the DSP is that it has far fewer and less sophisticated addressing
modes. Finally, DSPs frequently utilize modifications of the C language to more
efficiently express the signal processing parallelism and fractional arithmetic, and
thus maximize their speed. As a result, the DSP is much more efficient at signal
processing but less capable to accommodate the software associated with the
network protocols.
FPGAs have recently become capable of providing tremendous amounts of
multiply accumulate operations on a single chip, surpassing DSPs by more than
6
The mops take into account mathematical operations required to perform an algorithm, but not
the operations to calculate an effective memory address index, or offset, nor the operations to
perform loop counting, overflow management, or other conditional branching.
12
History and Background of Cognitive Radio Technology
an order of magnitude. By defining the on-chip interconnect of many gates, more
than 100 multiply accumulators can be arranged to perform multiply accumulate
processes at frequencies of more than 200 MHz. In addition to the DSP, FPGAs
can also provide the timing logic to synthesize clocks, baud rate, chip rate, time
slot, and frame timing, thus leading to a reasonably compact waveform implementation. By expressing all of the signal processing as a set of register transfer
operations and multiply accumulate engines, very complex waveforms can be
implemented in one chip. Similarly, complex signal processes that are not efficiently implemented on a DSP, such as Cordic operations, log magnitude operations, and difference magnitude operations, can all have the specialized hardware
implementations required for a waveform when implemented in FPGAs.
The downside of using FPGA processors is that the waveform signal processing is not defined in traditional software languages such as C, but in VHDL, a language for defining hardware architecture and functionality. The radio waveform
description in very high-speed integrated circuit (VHSIC) Hardware Design
Language (VHDL), although portable, is not a sequence of instructions and therefore
not the usual software development paradigm. At least two companies are working
on new software development tools that can produce the required VHDL, somewhat hiding this language complexity from the waveform developer. In addition,
FPGA implementations tend to be higher power and more costly than DSP chips.
All three of these computational resources demand significant off-chip memory. For example, a GPP may have more than 128 Mbytes of off-chip instruction
memory to support a complex suite of transaction protocols for today’s telephony
standards.
Today’s SDRs provide a reasonable mix of these computational alternatives to
assure that a wide variety of desirable applications can in fact be implemented at
an acceptable resource level.
In today’s SDRs, dedicated-purpose application-specific integrated circuit
(ASIC) chips are avoided because the signal processing resources cannot be
reprogrammed to implement new waveform functionality.
1.4.3 The Software Architecture of an SDR
The objective of the software architecture in an SDR is to place waveforms and
applications onto a software-based radio platform in a standardized way. These
waveforms and applications are installed, used, and replaced by other applications
as required to achieve the user’s objectives. To make the waveform and application interfaces standardized, it is necessary to make the hardware platform present
13
Chapter 1
a set of highly standardized interfaces. This way, vendors can develop their waveforms independent of the knowledge of the underlying hardware. Similarly, hardware developers can develop a radio with standardized interfaces, which can
subsequently be expected to run a wide variety of waveforms from standardized
libraries. This way, the waveform development proceeds by assuming a standardized set of interfaces (APIs) for the radio hardware, and the radio hardware translates commands and status messages crossing those interfaces to the unique
underlying hardware through a set of common drivers.
In addition, the method by which a waveform is installed into a radio, activated, deactivated, and de-installed, and the way in which radios use the standard
interfaces must be standardized so that waveforms are reasonably portable to
more than one hardware platform implementation.
According to Christensen et al., “The use of published interfaces and industry
standards in SDR implementations will shift development paradigms away from
proprietary tightly coupled hardware software solutions” [17]. To achieve this, the
SDR radio is decomposed into a stack of hardware and software functions, with
open standard interfaces. As was shown in Figure 1.3, the stack starts with the
hardware and the one or more data buses that move information among the various
processors. On top of the hardware, several standardized layers of software are
installed. This includes the boot loader, the operating system (OS); the board support package (BSP, which consists of input/output drivers that know how to control each interface); and a layer called the hardware abstraction layer (HAL). The
HAL provides a method for GPPs to communicate with DSPs and FPGA processors.
The US Government has defined a standardized software architecture, known
as the Software Communication Architecture (SCA), which has also been adopted
by defense contractors of many other countries worldwide. The SCA is a core
framework to provide a standardized process for identifying the available computational resources of the radio, matching those resources to the required resources
for an application. The SCA is built upon a standard set of OS features called
POSIX,7 which also has standardized APIs to perform OS functions such as file
management and computational thread/task scheduling.
The SCA core framework is the inheritance structure of the open application
layer interfaces and services, and provides an abstraction of underlying software
and hardware layers. The SCA also specifies a Common Object Request Broker
7
POSIX is the collective name of a family of related standards specified by the IEEE to define the
API for software compatible with variants of the Unix OS. POSIX stand for portable operating
system interface, with the X signifying the Unix heritage of the API [18].
14
History and Background of Cognitive Radio Technology
Architecture (CORBA) middleware, which is used to provide a standardized
method for software objects to communicate with each other, regardless of which
processor they have been installed on (think of it as a software data bus). The
SCA also provides a standardized method of defining the requirements for each
application, performed in eXtensible Markup Language (XML). The XML is parsed
and helps to determine how to distribute and install the software objects. In summary,
the core framework provides a means to configure and query distributed software
objects, and in the case of SDR, these will be waveforms and other applications.
These applications will have many reasons to interact with the Internet as well
as many local networks; therefore, it is also common to provide a collection of
standardized radio services, network services, and security services, so that each
application does not need to have its own copy of Internet Protocol, and other
commonly used functions.
1.4.4 Java Reflection in a Cognitive Radio
Cognitive radios need to be able to tell other cognitive radios what they are
observing that may affect the performance of the radio communication channel.
The receiver can measure signal properties and can even estimate what the transmitter meant to send, but it also needs to be able to tell the transmitter how to
change its waveform in ways that will suppress interference. In other words, the
cognitive radio receiver needs to convert this information into a transmitted message to send back to the transmitter.
Figure 1.7 presents a basic diagram for understanding cognitive radios. In this
figure, the receiver (radio 2) can use Java reflection to ask questions about the
internal parameters inside the receive modem, which might be useful to understand link performance. Measurements common in the design of a receiver, such
as the signal-to-noise ratio (SNR), frequency offset, timing offset, or equalizer
taps, can be read by the Java reflection. By examining these radio properties, the
receiver can determine what change at the transmitter (radio 1) will improve the
most important objective of the communication (such as saving battery life). From
that Java reflection, the receiver formulates a message onto the reverse link, multiplexes it into the channel, and observes whether the transmitter making that
change results in an improvement in link performance.
1.4.5 Smart Antennas in a Cognitive Radio
Current radio architectures are exploring the uses of many types of advanced
antenna concepts. A smart radio needs to be able to tell what type of antenna is
15
Chapter 1
Radio 1
Radio 2
Query/Reply
Query/Reply
Application
Reflection
Monitor
JTP
Reasoning
Agent
W Extractor
Application
Query
Query
Data Link
Reflection
Data Link
Monitor
Notify
Java Native Interface
JTP
Reasoning
Agent
Physical Layer
W Extractor
Notify
Java Native Interface
Physical Layer
Figure 1.7: Java reflection allows the receiver to examine the state variables of the transmit
and receive modem, thereby allowing the cognitive radio to understand what the communications channel is doing to the transmitted signal [19]. Copyright 2003 SDR Forum.
available, and to make full use of its capabilities. Likewise, a smart antenna
should be able to tell a smart radio what its capabilities are.
Smart antennas are particularly important to cognitive radio, in that certain
functionalities can provide very significant amounts of measurable performance
enhancement. As detailed in Chapter 5, if we can reduce transmit power, and
thereby allow transmitters to be closer together on the same frequency, we can
reduce the geographic area dominated by the transmitter, and thus improve the
overall spectral efficiency metric of MHzkm2.
A smart transmit antenna can form a beam to focus transmitted energy in the
direction of the intended receiver. At frequencies of current telecommunication
equipment in the range of 800–1800 MHz, practical antennas can easily provide
6–9 dB of gain toward the intended receiver. This same beamforming reduces the
energy transmitted in other directions, thereby improving the usability of the same
frequency in those directions.
A radio receiver may also be equipped with a smart antenna for receiving.
A smart receive antenna can synthesize a main lobe in the desired direction of the
intended transmitter, as well as synthesize a deep null in the direction of interfering transmitters. It is not uncommon for a practical smart antenna to be able to
synthesize a 20 dB null to suppress interference. This amount of interference suppression has much more impact on the MHzkm2 metric than being able to
transmit 20 dB more transmit power.
The utility of the smart antenna at allowing other radio transmitters to be
located nearby is illustrated in Figure 1.8.
16
History and Background of Cognitive Radio Technology
Figure 1.8: Utility of smart antennas.
A smart antenna allows a transmitter
(T) to focus its energy toward the
intended receiver (R), and allows a
receiver to suppress interference from
nearby interfering transmitters.
Transmitter Forms Beam Toward Intended Recipient
Receiver Forms Null Toward Interference Sources
R1,R4
T2
T1
Interfering
Signal Placed
in Null
R3
R 3,T 4
T3
1.5 Spectrum Management
The immediate interest to regulators in fielding cognitive radios is to provide new
capabilities that support new methods and mechanisms for spectrum access and
utilization now under consideration by international spectrum regulatory bodies.
These new methodologies recognize that fixed assignment of a frequency to one
purpose across huge geographic regions (often across entire countries) is quite
inefficient. Today, this type of frequency assignment results in severe underutilization of the precious and bounded spectrum resource. The Federal Communications
Commission (FCC; for commercial applications) and the National Telecommunications and Information Administration (NTIA; for federal applications) in the
United States, as well as corresponding regulatory bodies of many other countries,
are exploring the question of whether better spectrum utilization could be achieved
given some intelligence in the radio and in the network infrastructure.
This interest also has led to developing new methods to manage spectrum
access in which the regulator is not required to micromanage every application,
every power level, antenna height, and waveform design. Indeed, the goal of minimizing interference with other systems with other purposes may be reasonably
automated by the cognitive radio. With a cognitive radio, the regulator could
define policies at a higher level, and expect the equipment and the infrastructure to
resolve the details within well-defined practical boundary conditions such as
available frequency, power, waveform, geography, and equipment capabilities.
17
Chapter 1
In addition, the radio is expected to utilize whatever etiquette or protocol defines
cooperative performance for network membership.
In the United States, which has several broad classes of service, the FCC has
held meetings with license holders, who have various objectives. There are license
holders who retain their specific spectrum for public safety and for other such
public purposes such as broadcast of AM, FM, and TV. There are license holders
who purchased spectrum specifically for commercial telecommunications purposes. There are license holders for industrial applications, as well as those for
special interests.
Many frequencies are allocated to more than one purpose. An example of this
is a frequency allocated for remote control purposes—many garage door opener
companies and automobile door lock companies have developed and deployed
large quantities of products using these remote control frequencies. In addition,
there are broad chunks of spectrum for which NTIA has defined frequency and
waveform usage, and how the defense community will use spectrum in a process
similar to that used by the FCC for commercial purposes.
Finally, there are spectrum commons and unlicensed blocks. In these frequencies, there is overlapping purpose among multiple users, waveforms, and geography. An example of spectrum commons is the 2.4 GHz band, discussed in Section
1.5.1. The following sections touch on new methods for spectrum management,
and how they lead to spectrum efficiency.
1.5.1 Managing Unlicensed Spectrum
The 2.4 and 5 GHz band are popularly used for wireless computer networking.
These bands, and others, are known as the industrial, scientific, medical (ISM)
bands. Energy from microwave ovens falls in the 2.4 GHz band. Consequently, it
is impractical to license that band for a particular purpose. However, WiFi®
(802.11) and Bluetooth applications are specifically designed to coexist with a
variety of interference waveforms commonly found in this band as well as with
each other. Various types of equipment utilize a protocol to determine which frequencies or time slots to use and keep trying until they find a usable channel.
They also acknowledge correct receipt of transmissions, retransmitting data packets when collisions cause uncorrectable bit errors.
Although radio communication equipment and applications defined in these
bands may be unlicensed, they are restricted to specific guidelines about what frequencies are used and what effective isotropic radiated power (EIRP) is allowed.
Furthermore, they must accept any existing interference (such as that from
18
History and Background of Cognitive Radio Technology
microwave ovens and diathermy machines), and they must not interfere with any
applications outside this band.
Bluetooth and 802.11 both use waveforms and carrier frequencies that keep
their emissions inside the 2.4 GHz band. Both use methods of hopping to frequencies that successfully communicate and to error correct bits or packets that are
corrupted by interference. Details of Bluetooth and 802.11 waveform properties
are shown in Table 1.1.
The 802.11 waveform can successfully avoid interference from microwave
ovens because each packet is of sufficiently short duration that a packet can be
delivered at a frequency or during a time period while the interference is minimal.
Bluetooth waveforms are designed to hop to many different frequencies very rapidly, and consequently the probability of collision with a strong 802.11 or microwave is relatively small and correctable with error correcting codes.
The regulation of the 2.4 and 5 GHz bands consists of setting the spectrum
boundaries, defining specific carrier frequencies that all equipment is to use, and
limiting the EIRP. As shown in Table 1.1, the maximum EIRP is 1 W or less for
most of the wireless network products, except for the metropolitan WiMAX service, and the FCC-type acceptance is based on the manufacturer demonstrating
EIRP and frequency compliance.
It is of particular interest to note that each country sets its own spectral and
EIRP rules with regard to these bands. Japan and Europe each have regulatory
rules for these bands that are different from those of the United States. Consequently,
manufacturers may either (a) make three models, (b) make one model with a
switch to select to which country the product will be sold, (c) make a model that
is commonly compliant to all regional requirements, or (d) make a model that is
capable of determining its current location and implement the local applicable
rules. Method (d) is an early application of cognitive techniques.
1.5.2 Noise Aggregation
Communication planners worry that the combined noise from many transmitters
may add together and thereby increase the noise floor at the receiver of an important message, perhaps an emergency message. It is well understood that noise
power sums together at a receiver. If a receiver antenna is able to see the emissions of many transmitters on the same desired frequency and time slot, increasing
the noise floor will reduce the quality of the signal at the demodulator, in turn
increasing the bit error rate, and possibly rendering the signal useless. If the interfering transmitters are all located on the ground in an urban area, the interference
power from these transmitters decays approximately as the reciprocal of r3.8
19
Table 1.1: Properties of 802.11, Bluetooth, ZigBee, and WiMAX waveforms.
Name/description
WiFi; WLAN
Carrier frequency
5 GHz, 12 channels
(8 indoor, 4 point to
point)
802.11b
WiFi; WLAN
2.4 GHz, 3 channels
802.15.1
Bluetooth; WPAN
802.15.4
ZigBee; WPAN
802.16
WiMax, Wibro;
WMAN
2.4–2.4835 GHz, 79
channels each 1 MHz
wide, adaptive
frequency hopping at
1600 Hops/s
868.3 MHz, 1 channel;
902–928 MHz,
10 channels with
2 MHz spacing;
2405–2483.5 MHz,
16 channels with
5 MHz spacing
2–11 GHz (802.16a),
10–66 GHz (802.16);
BW 1.25, 5, 10,
and 20 MHz
20
Standard
802.11a
(802.11g both 802.11a
and 802.11b)
Modulation
52 carriers of OFDM, 48 data,
4 pilot, BPSK, QPSK, 16
QAM, or 64 QAM, carrier
separation 0.3125 MHz,
symbol duration 4 s,
with cyclic prefix 0.8 s,
Vitterbi R 1/2, 2/3, 3/4
CCK, DBPSK, DQPSK with
DSSS
GFSK (deviation 140–175 KHz)
32 chip symbols for 16 ary
orthogonal modulation with
OQPSK spreading at 2.0 Mcps
(2.4 GHz); DBPSK with BPSK
spreading at 300 Kcps
(868 MHz) or 600 Kcps
(915 MHz)
OFDM, SOFDM: 2048,
1024, 512, 256, and 128 FFT;
carriers, each QPSK, 16 QAM,
or 64 QAM; symbol rate 102.9 s
Data rate
Tx Pwr; EIRP
54, 48, 36, 27, 12–30 dBm
24, 1, 8, 12,
9, and 6 Mbps
11, 5.5, and
1.0 Mbps
57.6, 432.6,
and 721 Kbps,
2.1 Mbps
12–30 dBm
250 Kbps at
2.4 GHz; 40
and 20 Kbps
at 868 and
915 MHz
3–10 dBm
70 Mbps
40 dBm; EIRP 57.3 dBm
0–20 dBm
History and Background of Cognitive Radio Technology
(a detailed explanation of the exponents of range, r, is found in the Appendix to
Chapter 5). The total noise received is the sum of the powers of all such interfering transmitters. Even transmitters whose received power level is below the noise
floor also add to the noise floor. However, signals whose power level is extremely
small compared to the noise floor have little impact on the noise floor. If there are
100 signals each 20 dB below the noise, then that noise power will sum equal to
the noise, and raise the total noise floor by 3 dB. Similarly, if there are 1000 transmitters, each 30 dB below the noise floor, they can raise the noise floor by 30 dB
(see the Appendix to Chapter 5 for an explanation). However, the additional noise
is usually dominated by the one or two interfering transmitters that are closest to
the receiver.
In addition, we must consider the significant effect of personal communication
devices, which are becoming ubiquitous. In fact, one person may have several
devices all at close range to each other. Cognitive radios will be the solution to
this spectral noise and spectral crowding, and will evolve to the point of deployed
science just in time to help with the aggregated noise problems of many personal
devices all attempting to communicate in proximity to each other.
1.5.3 Aggregating Spectrum Demand and Use of Subleasing Methods
Many applications for wireless service operate with their own individual licensed
spectra. It is rare that each service is fully consuming its available spectrum.
Studies show that spectrum occupancy seems to peak at about 14 percent, except
under emergency conditions, where occupancy can reach 100 percent for brief
periods of time. Each of these services does not wish to separately invest in their
own unique infrastructure. Consequently, it is very practical to aggregate these
spectral assignments to serve a user community with a combined system. The
industry refers to a collection of services of this type as a trunked radio. Trunked
radio base stations have the ability to listen to many input frequencies. When a
user begins to transmit, the base station assigns an input and an output frequency
for the message and notifies all members of the community to listen on the
repeater downlink frequency for the message. Trunking aggregates the available
spectrum of multiple users and is therefore able to deliver a higher quality of
service while reducing infrastructure costs to each set of users and reducing the
total amount of spectrum required to serve the community.
Both public safety and public telephony services benefit from aggregating
spectrum and experience fluctuating demands, so each could benefit from the ability to borrow spectrum from the other. This is a much more complex situation,
21
Chapter 1
however. Public safety system operators must be absolutely certain that they can
get all the spectrum capacity they need if an emergency arises. Similarly, they
might be able to appreciate the revenue stream from selling access to their spectrum to commercial users who have need of access during times when no emergency conditions exist.
1.5.4 Priority Access8
If agreements can be negotiated between spectrum license holders and spectrum
users who have occasional peak capacity needs, it is possible to define protocols
to request access, grant access, and withdraw access. Thus, an emergency public
service can temporarily grant access to its spectrum in exchange for monetary
compensation. Should an emergency arise, the emergency public service can withdraw its grant to access, thereby taking over priority service.
In a similar fashion, various classes of users can each contend for spectrum
access, with higher-priority users being granted access before other users. This
might be relevant, for example, if police, fire, or military users need to use the cellular infrastructure during an emergency. Their communications equipment can
indicate their priority to the communications infrastructure, which may in turn
grant access for these highest-priority users first.9
By extension, a wide variety of grades of service for commercial users may
also prioritize sharing of commercially licensed spectrum. Users who are willing
to pay the most may get high priority for higher data rates for their data packets.
The users who pay least would get service only when no other grades of service
are consuming the available bandwidth.
1.6 US Government Roles in Cognitive Radio
1.6.1 DARPA
Paul Kolodzy was a Program Manager at DARPA when he issued a Broad Area
Announcement (BAA) calling for an industry day on the NeXt Generation (XG)
program to explore how XG communications could not only make a significant
impact on spectral efficiency of defense communications, but also significantly
reduce the complexity of defining the spectrum allocation for each defense user.
8
Cellular systems already support priority access; however, there is reported to be little control
over the allocation of priority or the enforcement process.
9
This technique is implemented in code division multiple access (CDMA) cellular communications.
22
History and Background of Cognitive Radio Technology
Shortly after proposals were sent in to DARPA, Kolodzy moved to the FCC to
further explore this question and Preston Marshall became the DARPA program
manager. Under Marshall’s XG program, several contractors demonstrated that a
cognitive radio could achieve substantial spectral efficiency in a non-interfering
method, and that the spectrum allocation process could be simplified. Basic principles of spectrum efficiency are discussed by Marshall in Chapter 5.
In the same time frame, Jonathan Smith worked as a DARPA project manager
to develop intelligent network protocols that could learn and adapt to the properties of wireless channels to optimize performance under current conditions. Some
of this work is discussed in Chapter 9.
1.6.2 FCC
On May 19, 2003, the FCC held a hearing to obtain industry comments on cognitive radio. Participants from the communications industry, radio and TV broadcasters, public safety officers, telecommunications systems operators, and public
advocacy participants all discussed how this technology might interact with the
existing spectrum regulatory process. Numerous public meetings were held subsequently to discuss the mechanics of such systems, and their impact on existing
license holders of spectrum (see Section 1.4 and Chapter 2). The FCC has been
actively engaged with industry and very interested in leveraging this technology.
1.6.3 NSF/CSTB Study
President George W. Bush has established the Spectrum Policy Task Force (SPTF)
to further study the economic and political considerations and impacts of spectrum policy. In addition to the FCC’s public meetings, the National Science
Foundation (NSF) also has held meetings on the impact of new technologies to
improve spectrum efficiency. A Committee chaired by Dale Hatfield and Paul
Kolodzy heard testimony from numerous representatives, leading to the SPTF
report further described in Chapter 2.
The Computer Science and Telecommunications Board (CSTB) is a specific
work group of the NSF. This work group produces books and workshops on
important topics in telecommunications. CSTB held numerous workshops on the
topic of spectrum management since its opening meeting at the FCC in May 2003.
These meetings have resulted in reports to the FCC on various cognitive radio topics. Recently, a workshop was held on the topic of “Improving Spectrum Management through Economic and Other Incentives.” This activity has been guided by
Dale Hatfield (formerly Chief of the Office of Science and Technology at the FCC,
23
Chapter 1
now Adjunct Professor at University of Colorado); William Lehr (Economist and
Research Associate at Center for Technology Policy and Industrial Development
at Massachusetts Institute of Technology); and Jon Peha (associate director of the
Center for Wireless and Broadband Networking at Carnegie Mellon University).
1.7 How Smart Is Useful?
The cognitive radio is able to provide a wide variety of intelligent behaviors. It
can monitor the spectrum and choose frequencies that minimize interference to
existing communication activity. When doing so, it will follow a set of rules that
define what frequencies may be considered, what waveforms may be used, what
power levels may be used for transmission, and so forth. It may also be given
rules about the access protocols by which spectrum access is negotiated with
spectrum license holders, if any, and the etiquettes by which it must check with
other users of the spectrum to ensure that no user hidden from the node wishing
to transmit is already communicating.
In addition to the spectrum optimization level, the cognitive radio may have
the ability to optimize a waveform to one or many criteria. For example, the radio
may be able to optimize for data rate, for packet success rate, for service cost, for
battery power minimization, or for some mixture of several criteria. The user does
not see these levels of sophisticated channel analysis and optimization except as
the recipient of excellent service.
The cognitive radio may also exhibit behaviors that are more directly apparent
to the user. These behaviors may include: (a) awareness of geographic location,
(b) awareness of local networks and their available services, (c) awareness of the
user and the user’s biometric authentication to validate financial transactions, and
(d) awareness of the user and his or her prioritized objectives. This book explores
each of these technologies. Many of these services will be immediately valuable
to the user without the need for complex menu screens, activation sequences, or
preference setup processes.
The cognitive radio developer must use caution to avoid adding cognitive
functionality that reduces the efficiency of the user at his or her primary tasks. If
the user thinks of the radio as a cell phone and does not wish to access other networks, the cognitive radio developer must provide a design that is friendly to the
user, timely and responsive, but is not continually intruding with attempts to be
helpful by connecting to networks that the user does not need or want. If the
radio’s owner is a power user, however, the radio may be asked to watch for
multiple opportunities: access to other wireless networks for data services,
24
History and Background of Cognitive Radio Technology
notification of critical turning points to aid navigation, or timely financial information, as a few simple examples.
One of the remaining issues in sophisticated software design is a method for
determining whether the cognitive services the radio might offer will be useful.
Will the services be accomplished in a timely fashion? Will the attempted services
be undesired and disruptive? Will the services take too long to implement and
arrive too late to be usable? The cognitive radio must offer functionality that is
timely and useful to its owner, and yet not disruptive. Like “Radar” O’Reilly, we
want the cognitive radio to offer support of the right type at the right time, properly prioritized to the user needs given sophisticated awareness of the local situation, and not offering frequent useless or obvious recommendations. We will
explore this topic in Chapter 16.
1.8 Organization of this Book
In Chapter 2, Paul Kolodzy describes the regulatory policy motivations, activities,
and initiatives within US and international regulatory bodies to achieve enhanced
spectral efficiency. Chapter 3, by Max Robert of Virginia Tech and Bruce Fette of
General Dynamics C4 Systems, describes the details of hardware and software
architecture of SDRs, and explains why an SDR is the primary choice as the basis
for cognitive radios. Chapter 4, by John Polson of Bell Helicopter, is about the
technologies required to implement basic services in a cognitive radio.
In Chapter 5, Preston Marshall of DARPA deals with spectrum efficiency and
the DARPA programs that demonstrated the feasibility of cognitive radio principles. Chapter 6, by Robert Wellington of the University of Minnesota, introduces
the cognitive policy engine. The policy engine provides an efficient mechanism to
express the rules applied to the function of the cognitive radio. This includes regulatory policy, network operator policy, radio equipment capability, and real-time
checking that, at the end of all the cognitive logic, the radio’s planned performance is allowed within its rules.
Chapter 7, by Thomas Rondeau and Charles Bostian of Virginia Tech, provides
a detailed analysis of cognitive techniques at the physical and medium access
control layers. These techniques are discussed in the context of genetic algorithms that can adapt multiple waveform properties in order to optimize link
performance.
In Chapter 8, John Polson and Bruce Fette describe a wide variety of methods
by which a radio can determine its local position, and thereby to use time and location information to assist the network or the user. In Chapter 9, Jonathan Smith of
25
Chapter 1
DARPA10 covers cognitive techniques in network adaption. This technology
allows the radio to be aware of local networks and their properties and services.
Smith focuses on how networks apply intelligence to the selection of network protocols in order to optimize network performance in spite of the differences
between wireless and wired systems.
Chapter 10, by Joe Campbell, Bill Campbell, Scott Lewandowski, and Cliff
Weinstein of Lincoln Laboratory, is about using speech as an input/output mechanism for the user to request and access services, as well as to authenticate the user
to the radio network and services. Speech analysis tools extract basic properties of
speech. These properties are further analyzed in different ways to result in word
recognition, language recognition, or speaker identification.
Chapter 11, by Youping Zhao, Bin Le, and Jeffrey Reed of Virginia Tech,
deals with the ways in which network infrastructure can provide cognitive functionality and support services to the user, even if the subscriber’s unit is of low
power or small computational capability.
Chapter 12, by Vince Kovarik of Harris Corporation, provides extensive coverage of learning technologies and techniques, and how these are applied to the
cognitive radio application.
In Chapter 13, Mitch Kokar, David Brady, and Kenneth Baclawski of
Northeastern University give a detailed overview of how to represent the types of
knowledge a radio would need to know in order to behave intelligently. This chapter deals with the storage, and analysis of this information as additional spatial,
temporal, radio, network, and application data are accumulated.
In Chapter 14, Joseph Mitola III of the MITRE Corporation describes how to
develop a complete radio and how to make the various radio modules work with
each other as an integrated cognitive system.
In Chapter 15, Jody Neel, Jeff Reed, and Allen McKenzie of Virginia Tech
provide a detailed analysis of game theory and how it is used to model the performance choices and system-level behaviors of networks consisting of a mixture
of cognitive and non-cognitive radios.
Chapter 16 completes this book with an overview of the remaining key problems that must be solved to allow cognitive radio to fulfill its considerable potential and become a recognized and successful self-sustaining industry.
10
Jonathan Smith is currently on the faculty of the University of Pennsylvania, PA.
26
History and Background of Cognitive Radio Technology
References
[1] A. Oppenheim and R. Schaefer, Discrete Time Signal Processing, Prentice Hall,
Englewood Cliffs, NJ, 1989.
[2] L. Rabiner and R. Schaefer, Digital Processing of Speech Signals, Prentice Hall,
Englewood Cliffs, NJ, 1978.
[3] L.R. Rabiner, J.H. McClellan and T.W. Parks, “FIR Digital Filter Design
Techniques Using Weighted Chebyshev Approximations,” Proceedings of the
IEEE, Vol. 63, 1975, pp. 595–610.
[4] T.W. Parks and J.J. McClellan, “Chebyshev Approximation for Nonrecursive
Digital Filters with Linear Phase,” IEEE Transactions on Circuit Theory, Vol. 19,
1972, pp. 189–194.
[5] J. Flanagan, Speech Synthesis and Perception, Springer Verlag, New York, 1972.
[6] Fred Harris, Multirate Signal Processing for Communication Systems, Prentice
Hall, Englewood Cliffs, NJ, 2004.
[7] http://www.mathworks.com/company/aboutus/founders/jacklittle.html
[8] J. Markel and A. Gray, Linear Prediction of Speech, Springer Verlag, Berlin,
Heidelburg and New York, 1976.
[9] http://www.intel.com/pressroom/kits/bios/moore.htm
[10] http://www.dspguide.com/filters.htm
[11] F. Rosenblatt, “The Perceptron: A probabilistic Model for Information Storage and
Organization in the Brain,” Cornell Aeronautical Laboratory, Psychological
Review, Vol. 65, No. 6, 1958, pp. 386–408.
[12] J. Baker, Stochastic Modeling for Speech Recognition, Doctoral Thesis, Department
of Computer Science, Carnegie Mellon University, Pittsburgh, PA, 1976.
[13] K. Lee, R. Reddy and Hsiao-wuen, “An Overview of the SPHINX Speech
Recognition System,” IEEE Transactions on Acoustics Speech and Signal
Processing, pp. 34–45, January 1990.
[14] J.H. Reed, Software Radio: A Modern Approach to Radio Engineering, Prentice
Hall, Englewood Cliffs, NJ, 2002.
[15] http://csrc.ncsl.nist.gov/cryptval/des/des.txt
[16] http://csrc.nist.gov/CryptoToolkit/aes/
[17] E. Christensen, A. Miller and E. Wing, “Waveform Application Development
Process for Software Defined Radios,” IEEE Milcom Conference, pp. 231–235,
Oct. 22, 2000.
[18] http://en.wikipedia.org/wiki/POSIX
[19] J. Wang, “The Use of Ontologies for the Self-Awareness of Communication
Nodes,” in Proceedings of the SDR Forum Technical Conference, Orlando, FL,
November 2003.
27
This page intentionally left blank
CHAPTER 2
Communications Policy and
Spectrum Management
Paul Kolodzy
Kolodzy Consulting
Centreville, VA, USA
2.1 Introduction
New technologies impact the worlds of commerce and policy. This is especially
true of disruptive technologies that significantly alter either the realities or perceptions within these worlds. Cognitive radio technology has the potential of affecting the marketplace for radio devices and services as well as changing the means
by which wireless communications policy is developed and implemented. One of
the key parameters that must be addressed to enter the radio market is access to
radio spectrum. Once access is obtained, the capacity to manage interference
becomes a key attribute in order to increase the number of users. Throughput is
critical in order to maximize benefit (for the device) or maximize revenue (for the
service). Radio frequency (RF) spectrum access and interference management are
thus the primary roles of spectrum management.
Cognitive radio technology has the potential of being a disruptive force within
spectrum management. Spectrum management, since the dawn of radio technology, has been within the domain of management agencies, both private and government. Therefore, it has required a person-in-the-loop. The ability of a device to be
aware of its environment and to adapt to enhance its performance, and the performance of the network, allows a transition from a manual, oversight process to an
automated, device-oriented process. This ability has the potential to allow a much
more intensive use of the spectrum by lowering of spectrum access barrier to entry
29
Chapter 2
for new devices and services. It also has the potential to radically change how
policy should be developed in order to account for these new uses of the spectrum,
and it can fundamentally change the role of the spectrum policy-maker and regulator.
In this chapter, Section 2.2 discusses the cognitive radio technology enablers.
Section 2.3 addresses spectrum access and how cognitive radio needs various
types of policies, depending on the density of spectral activity and the types of
usages. This section also provides examples of spectral activity measurement.
Section 2.4 discusses the challenges to equipment developers associated with a
policy-based approach to spectrum management. Section 2.5 presents the challenges to the regulators to manage spectrum policy through radios and networks
that operate based on policy. Section 2.6 discusses the global interest and activity
in policy-based cognitive radio. Finally, Section 2.7 provides a summary of this
chapter’s major issues.
2.2 Cognitive Radio Technology Enablers
The development of wideband power amplifiers, synthesizers, and analog-to-digital
converters (ADCs) is providing a new class of radios: the software-defined radio
(SDR) and its software and cognitive radio cousins. Although at the early stages of
development, this new class of radio ushers in new possibilities as well as potential
pitfalls for technology policy. The flexibility provided by the cognitive radio class
of radios allows for more dynamics within radio operations. The same flexibility poses
challenges for certification and the associated liability through potential misuse.
SDRs provide software control of a variety of modulation techniques, wideband and narrowband operation, transmission security (TRANSEC) functions
(such as hopping), and waveform requirements. In essence, components can be
under digital control and thus defined by software. The advantage of an SDR is
that a single system can operate under multiple configurations, providing interoperability, bridging, and tailoring of the waveforms to meet the localized requirements. SDR technology and systems have been developed for the military. The
digital modular radio (DMR) system was one of the first SDR systems. Recently
the US Defense Advanced Research Projects Agency (DARPA) developed the
Small Unit Operations Situational Awareness Systems (SUO SAS), which was a
man-portable SDR operating from 20 MHz to 2.5 GHz. The level of success of
these programs has led to the Joint Tactical Radio System (JTRS) initiative to
develop and procure SDR systems throughout the US military.
SDRs exhibit software control over a variety of modulation techniques and
waveforms. Software radios (SRs) specifically implement the waveform signal
30
Communications Policy and Spectrum Management
processing in software. This additional caveat essentially has the radio being
constructed with a RF front-end, a down-converter to an intermediate frequency
(IF) or baseband, an ADC, and then a processor. The processing capacity therefore limits the complexity of the waveforms that can be accommodated.
A cognitive radio adds both a sensing and an adaptation element to the software defined and software radios. Four new capabilities embodied in cognitive
radios will help enable dynamic use of the spectrum: flexibility, agility, RF sensing, and networking [1].
●
●
●
●
Flexibility is the ability to change the waveform and the configuration of a device.
An example is a cell tower that can operate in the cell band for telephony purposes but change its waveform to get telemetry from vending machines during
low usage, or other equally useful, schedulable, off-peak activity. The same band
is used for two very different roles, and the radio characteristics must reflect the
different requirements, such as data rate, range, latency, and packet error rate.
Agility is the ability to change the spectral band in which a device will operate.
Cell phones have rudimentary agility because they can operate in two or more
bands (e.g., 900 and 1900 MHz). Combining both agility and flexibility is the
ultimate in “adaptive” radios because the radio can use different waveforms in
different bands. Specific technology limitations exist, however, to the agility and
flexibility that can be afforded by current technology. The time scale of these
adaptations is a function of the state of technology both in the components for
adaptation as well as the capacity to sense the state of the system. These are classically denoted as the observable/controllable requirements of control systems.
Sensing is the ability to observe the state of the system, which includes the
radio and, more importantly, the environment. It is the next logical component
in enabling dynamics. Sensing allows a radio to be self-aware, and thus it can
measure its environment and potentially measure its impact to its environment.
Sensing is necessary if a device is to change in operation due to location, state,
condition, or RF environment.
Networking is the ability to communicate between multiple nodes and thus
facilitate combining the sensing and control capacity of those nodes. Networking,
specifically wireless networking, enables group-wise interactions between radios.
Those interactions can be useful for sensing where the combination of many
measurements can provide a better understanding of the environment. They can
also be useful for adaptation where the group can determine a more optimal use
of the spectrum resource over an individual radio.
31
Chapter 2
●
●
●
●
●
SDR: “A radio that includes a transmitter in which the operating parameters of frequency range, modulation type or maximum output power (either radiated or conducted), or the circumstances under which the
transmitter operates can be altered by making a change in software without making any changes to
hardware components that affect the RF emissions.”—Derived from the US FCC’s Cognitive Radio
Report and Order, adopted 2005-03-10.
Cognitive radio: A radio or system that senses and is aware of its operational environment and can be
trained to dynamically and autonomously adjust its radio operating parameters accordingly.
[It should be noted that “cognitive” does not necessarily imply relying on software. For example, cordless
telephones (no software) have long been able to select the best authorized channel based on relative
channel availability.]
Policy-based radio: A radio that is governed by a predetermined set of rules for behavior. The rules
define the operating limits of such a radio. These rules can be defined and implemented:
– During manufacture
– During configuration of a device by the user
– During over-the-air provisioning and/or
– By over-the-air control.
Software reconfigurable radio: A SDR that: (1) incorporates software-controlled antenna filters to dynamically select receivable frequencies, and (2) is capable of downloading and installing updated software
for controlling operational characteristics and antenna filters without manual intervention.
DFS:
1. “A general term used to describe mitigation techniques that allow, amongst others, detection, and
avoidance of co-channel interference with other radios in the same system or with respect to other
systems.”—From current version of WP8A PDNR on SDR;
2. “The ability to sense signals from other nearby transmitters in an effort to choose an optimum
operating environment.”—Derived from the US FCC’s Cognitive Radio Report and Order, adopted
2005-03-10.
Figure 2.1: ITU GSC proposed definitions
These new technologies and radio classes, albeit in their nascent stages of
development, are providing many new tools to the system developer while allowing for more intensive use of the spectrum. However, an important characteristic
of each of these technologies is the ability to change configuration to meet new
requirements.1 This capacity to react to system dynamics will require the
development of new spectrum policies in order to take advantage of these new
characteristics.
The definitions for the various radio technologies are developed by the
International Telecommunication Union (ITU) with help from many of its
member organizations. Figure 2.1 provides a list of radio technology definitions
1
The question of how cognitive radios would apply physical layer adaption in sensor networks
that transmit and receive over very low duty cycles has not been adequately studied. Although it is
assumed that they could also benefit from cognitive radio adaption, the time to reach network stability is lengthened by the low duty cycles. It seems that if the spectral dynamics exceed the time
to reach stable performance, or consume more bandwidth for parametric exchange and adaption
parameters than the network can effectively provide, then such sensor networks may not be able to
fully benefit from cognitive radio techniques. However, if other systems operating in the same environment are cognitive radios with stable adaption strategies, then sensor networks may still benefit.
32
Communications Policy and Spectrum Management
proposed by the Global Standards Collaboration (GSC) group within the ITU.
Definitions for policy-based radios and dynamic frequency selection (DFS) radios
are also provided. These two new radio classes are specific implementations of
cognitive radios. Policy-based radios are discussed in Section 2.6 with an emphasis on how cognitive radio technology can impact the development and implementation of communications policy. DFS radios are addressed in Section 2.3.3. These
advanced radio technologies are enabling a multitude of new radio concepts, as
discussed in Section 2.3.2.
2.3 New Opportunities in Spectrum Access
Two general management methods allow access to the RF spectrum: spectrum
access licenses and unlicensed devices. Spectrum licenses are issued by the
appropriate regulatory agency within the nation. The licenses include a band, a
geographic region, and the allowable operational parameters (e.g., in-band and
out-of-band transmission levels). Although the licenses have a finite duration,
there is an expectation of renewal. There is also a level of protection from
interference from other systems accessing the spectrum. Unlicensed devices, also
called licensed-free devices and licensed-by-rule devices, are provided frequency
bands and transmission characteristics (albeit at much lower transmission power
levels), and are not provided regulatory protection from interference.
2.3.1 Current Spectrum Access Techniques
The RF spectrum is organized by allocations and assignments. Allocations determine the type of use and the respective transmission parameters. The allocations
are specified by each individual nation determined by sovereign needs and international agreements. International agreements can be bilateral or multilateral
treaties or resolutions by the ITU during the periodic World Radio Communication
Conferences (WRCs). Figures 2.2 and 2.3 are examples of spectrum allocations
for the United States and New Zealand, respectively. Each color is an indication of
a service type that is allocated to that frequency band across the entire nation.
Many of the primary allocations (television (TV), frequency modulation (FM)
radio, global positioning systems (GPS), etc.) are identical. The vertical splitting
that occurs in some bands depicts the instance of multiple allocations (primary,
co-primary, and secondary) within the band. The intensity of usage in the United
States is much higher than in New Zealand. This is really noticeable in the figures
by the large number of multiple allocations needed to provide enough capacity for
the numerous wireless services for commercial and noncommercial applications,
such as defense, air traffic, and scientific exploration.
33
Chapter 2
Figure 2.2: US spectrum allocations [2].
Assignments are individual licenses within an allocation. Assignments are
provided by the method of choice of the sovereign state. Assignments can also be
limited by geographic extent. There are many ways to obtain assignments, as depicted
by Figure 2.4. Currently, three basic types of assignment methods—command and
control, auctions, and protocols and etiquettes are generally employed.
●
Command and control: Command and control assignments are provided by the
regulatory agency by reviewing specific licensing applications and choosing the
prospective licensee by criteria specific to the national goals. Detractors of this
assignment technique call it a “beauty contest” because only a handful of
regulators determines the relative value between potential services and license
holders. Since the US Radio Act of 1934, the Federal Communications
Commission (FCC) has had the regulatory authority to decide which firms
34
Communications Policy and Spectrum Management
Figure 2.3: New Zealand spectrum allocations [3].
should get licenses in order to bring spectrum to its highest and best use. The
FCC held public hearings to gather information to make these determinations.
In the early 1980s the number of applications for licenses was growing so large
that the command and control system ground to a halt. The FCC, to enable
more capacity for the command and control assignment mechanism, decided to
start awarding licenses by lottery. The assignment process was still in complete
control of the regulators; however, there were no restrictions as to who could
participate in the lottery. One year, for instance, a group of dentists won a
license to run cellular phones on Cape Cod, and then promptly sold it to
Southwestern Bell for $41 million. This set of events disturbed regulators
because the spectrum assignment process became an investment opportunity
rather than providing a service (albeit for profit). This directly challenged their
authority in using spectrum for the benefit of the country (i.e., its citizens).
●
Auctions: Spectrum assignments through auctions are fairly new. New Zealand,
with the Radiocommunications Act 1989, enabled the use of market-driven
35
Chapter 2
Third
THIRD
party
PARTY
Licenses
LICENSES
Command
and
control
Auctions
Band Managers
Market forces
dominates
Spectrum access
and coordination
using market
forces
Authoritative
entity selects
Secondary
markets
Market forces
with negotiations
Lottery
Device
DEVICE
centric
CENTRIC
Unlicensed
Device certification
Opportunistic Limited emission
Authoritative
determines
eligibility, random
drawing
Device certification
Highly complex
Sensing capability
Unlimited potential
levels
Figure 2.4: Spectrum access regimes.
allocation mechanisms for assignment of spectrum rights. Although the act did
not explicitly provide the market mechanisms, by 1996 the mechanisms for
selling spectrum by auction were in place. Specifically, the use of competitive
bidding was authorized in granting licenses to qualified applicants.2 In the
Telecommunications Act of 1996, the United States also enabled the use of
auctions for spectrum access licenses. The Wireless Telegraphy Act of 1998
enabled auctions in the United Kingdom. Auctions are now a common form for
spectrum assignments throughout the industrial world. The United States,
Australia, Mexico, Canada, New Zealand, and the European Union now employ
auctions regularly. The current spectrum allocations in the United States and
New Zealand presented in Figures 2.1 and 2.2, respectively, show the many
competing applications for spectrum.
2
Section 309 [47 U.S.C 309] Subsection J: “USE OF COMPETITIVE BIDDING. If mutually
exclusive applications are accepted for filing for any initial license or construction permit
which will involve a use of the electromagnetic spectrum described in paragraph (2), then the
Commission shall have the authority, subject to paragraph (10), to grant such license or permit to a
qualified applicant through the use of a system of competitive bidding that meets the requirements
of this subsection.”
36
Communications Policy and Spectrum Management
●
Protocols and etiquettes: Unlicensed devices and amateur licensees do not
have, per se, specific frequency assignments. The allocation allows these
devices to operate within a band, and the selection of a particular frequency
is accomplished through protocols and etiquettes. Protocols are explicit
interactions for spectrum access. Carrier Sense Multiple Access with Collision
Avoidance (CSMA/CA) is a protocol for media access and control that requires
information to be communicated between devices. Etiquettes are rules that are
followed without explicit interaction between devices. Simple etiquettes, such
as “listen before talk,” DFS, and power density limits are one-sided processes.
The amateur radio operators have constructed a well-tested and successful
manual set of etiquettes. For example, the developers of DX tuners (for longdistance transmission and reception by ham radio operators) have etiquettes such
as “If there are other users logged in, always ask in the chat window before you
tune the receiver!” Unlicensed devices are operated on a much shorter time
scale. Industry groups such as the Institute of Electrical and Electronics
Engineers (IEEE), form groups in order to develop standards to promote device
interoperability. The 802.3 (Ethernet) and 802.11 (wireless local area network,
or WLAN) are two such protocol standards. New etiquettes, such as DFS, have
been developed for the new 5–5.8 GHz 802.11a band (IEEE 802.11h).
Spectrum access for services provided by licenses is controlled directly by the
service provider. In the case of public broadcast services such as FM radio and
TV, the waveforms and/or protocols are defined by the rules provided by the
regulators. These rules, although provided by the regulators, are developed with
both the service providers and the device manufacturers. Being a broadcast service, the broadcast station controls access simply by its own use of the frequency
band. In the case of providers of two-way communications or private broadcast,
the waveforms and protocols are defined independently by the license holder.
However, market forces tend to drive groups of license holders to compatible
technology in order to increase either the interoperability with other license holders or to increase the geographic footprint of their service. A classic example of
this independent development of industry standards has been the commercial
mobile radio service (CMRS, or cellular service) and the development of the
capacity to roam from service provider to service provider. The development
of roaming has generally been considered to have been the threshold event in
making radiotelephony a viable service.
The operational envelopes for unlicensed devices are defined by rules in
three general groupings: unintentional, incidental, and intentional radiators. An
37
Chapter 2
unintentional radiator, per FCC definition, is a device that intentionally generates
RF energy for use within the device, or that sends RF signals by conduction to
associated equipment via connecting wiring, but which is not intended to emit
RF energy by radiation or induction. One example would be a TV receiver with
leakage from an oscillator in the receiver RF chains. Another very common example concerns a computer and display that typically exhibit harmonics up to more
than 3 GHz. The first example has the emission contained in a particular band and
the second example is a broadband, noise emission.
An incidental radiator, per FCC definition, is a device that generates RF
energy during the course of its operation, although the device is not intentionally
designed to generate or emit RF energy. Examples of incidental radiators are
direct current (DC) motors, mechanical light switches, or even the radiation from
a hair dryer.
In both the unintentional and incidental radiator cases, the regulatory agency
has set limits on the emission levels. The allowable emission levels are generally
very low due to the general and ubiquitous deployment of these devices. The
devices may be in extremely close proximity to licensed service devices. Usually,
the consumer does not know the potential impact of the device and thus cannot
easily resolve any interference issues. Therefore, the aggregated interference from
many such devices must remain small.
Intentional radiators are devices that are specifically constructed for a communications application. A prime example is that of baby monitors, a one-way communications system for use in the home. Other examples of intentional unlicensed
radiators are the data networks using the 802.11 standard. These devices are
allowed to radiate at higher levels than the unintentional radiators but still at low
overall power spectral density. The low power spectral density provides some
assurance that the interference potential to a primary license holder within that
band is low. The regulators determine the allowable transmission parameters.
There is no licensee, however, so the device manufacturer has the burden to conform to the transmission rules.
The primary difference between licensed and unlicensed device spectrum
access is the afforded interference protection. The unlicensed device always operates as a secondary user within a band and is allowed to operate only in a not-tointerfere basis, per the transmission rules. The unlicensed device is not provided
any interference assurances from either the primary use devices or other unlicensed devices. It is for this reason that device manufacturers and application
designers develop standard protocols and/or etiquettes in order to reduce the
potential of like-device interference. This is true with the development of the
38
Communications Policy and Spectrum Management
802.11 specifications for WLANs. These specifications are widely used with the
popular consumer products in the 2.4 and 5.15–5.825 GHz bands. However, unlike
devices such as 2.4 GHz cordless telephones and 802.11g devices, which operate
in the industrial, scientific, and medical (ISM) band, the expanded 802.11h
devices operating in the 5.15–5.825 GHz band have the potential to interfere with
incumbent devices. The common use of the 2.4 GHz ISM band is for microwave
ovens, which are not overly susceptible to radiated energy.
2.3.2 Opportunistic Spectrum Access
The current toolbox of spectrum access techniques is limited by the capacity of
both the devices and the policy that are in place. As was shown in the spectrum
allocations charts depicted in Figures 2.2 and 2.3, spectrum less than 6 GHz is
completely allocated for various services. However, the use of the assignments
in time and space is not complete. An example of a measurement of spectrum
utilization is shown in Figure 2.5. This time-frequency measurement of the
700–800 MHz band was made over an 18-hour period in Hoboken, New Jersey,
104
fMHz
02:17
10
Time (sec)
20:06
14:11
5
08:18
02:27
0
700
710
720
730
740
Average duty cycle 0.30793
750
fMHz
760
770
780
790
800
700
710
750
fMHz
760
770
780
790
800
20:40
Duty cycle
1
0.5
0
720
730
740
Figure 2.5: Spectrum utilization. Utilization in Hoboken, New Jersey, 700–800 MHz, made by
Shared Spectrum Corporation over an 18-hour period.
39
Chapter 2
which is directly across the Hudson River from Manhattan, New York. It
clearly indicates that there are portions of the spectrum that are continuously
accessed, portions that are never accessed, and portions that are accessed for a
fraction of the time. These “white spaces” in the spectrum have created interest
within the US Department of Defense (DoD) to begin research into more efficiently using spectrum. The FCC also saw the potential for more intensive use
of the spectrum as well as revisiting the age-old claim of scarcity of the RF
spectrum.
The onset of SDR and cognitive radio technology enables opportunistic
spectrum access (OSA). OSA looks for “holes” in the spectrum and then adjusts
the link parameters to conform to the hole. That is, the radio transmits over
sections of the spectrum that are not in use. However, it has the additional
complexity of listening for other transmitters in order to vacate a hole when other,
nonopportunistic spectrum devices are accessing it. This technology combines
flexible waveform capacity with sensing and adaptive frequency technology.
The potential gain is the higher utilization of infrequently used spectrum. It has
been estimated that on average, less than 5 percent, and possibly as little as
1 percent, of the spectrum less than 3 GHz as measured in frequency-space-time
is used.3 Opportunistic spectrum technology is under development by the US
DoD with the goal of increasing accessible spectrum by a factor of 20. The
European Union Information Society Technologies (IST) as well as the US
National Science Foundation (NSF) and other research organizations, as of
2005, are actively pursuing this technology. Figure 2.6 provides a chronology
of events in OSA activities.
The availability of SDR also allows high-priority users, such as in a public
service context, to access the radio spectrum on an “as-needed” basis. This is
called interruptible spectrum access because the normal user’s spectrum access is
interrupted in order to provide the spectral resource to a higher priority user (see
Figure 2.7). For example, in a major regional disaster there will be a significant
increase in public safety users due to the influx of responders from outside the
immediate area. Such an influx of radio users would require additional spectrum
3
This percentage is a function of where the measurement is made and what sensitivity thresholds
are used. Rural areas have an extremely low utilization. In the worst-case condition of downtown New York City during the Republican Convention, the utilization rose to approximately
20 percent. The peak utilization in San Diego on one sample day was 7 percent. Because the
numbers are location and emergency dependent, precise numbers cannot be stated universally,
but utilization is definitely low.
40
Communications Policy and Spectrum Management
●
●
●
●
●
●
●
●
1999–2000 (United States): Localized set of measurements conducted by the DARPA indicated that
spectrum use was not very high.
2002 (United States): DARPA initiates the XG (neXt Generation) project to investigate the potential for
the military to share spectrum spatially and temporally with multiple devices.
2002 (United States): The FCC Spectrum Policy Task Force (SPTF) concludes that spectrum access is
a more significant problem than spectrum scarcity. The SPTF recommends that new rules be developed
to allow more intensive access to the spectrum, including opportunistic spectrum.
2003 (European Union): IST includes DSA technologies as part of the sixth Framework Programme of
R&D.
2003 (United States): The NSF initiates research projects in spectrum measurements and DSA.
2004 (United States): The FCC issues a Notice of Proposed Rulemaking on Facilitating Opportunities for
Flexible, Efficient, and Reliable Spectrum Use Employing Cognitive Radio Technologies.
2004 (European Union): End-to-End Reconfigurability (E2R) Project initiated in IST.
2005 (United States): DARPA XG and NSF projects complete series of spectrum occupancy
measurements indicating less than 10 percent occupancy in time-space under 3 GHz.
Figure 2.6: Chronology of opportunistic spectrum.
Request additional
spectrum
Service A
Service B
Service C
Service A and C interrupted
and reduce spectral footprint
Service A
Service B
Service C
Service B increases
spectral footprint
Service A
Service B
Service C
Figure 2.7: Interruptible spectrum–
aperiodic increase in spectral
requirements (i.e., public safety) may
be addressed by interrupting part of
the band of other services. This can
be accomplished through central
coordination or beaconing.
to cope with the increased load. Additional spectrum could be temporarily
obtained from adjacent spectral bands in the immediate area of the disaster. After
the need diminished, perhaps in minutes or hours, the spectrum would be released
back to the primary licensee. Mechanisms may be developed to compensate primary licensees for their inconvenience or loss of revenue.
If policies can be malleable to specific conditions in order to allow greater
access to the RF spectrum, then the cost barriers for new entrants should be
reduced. Thus, new consumer products can be developed using lower-cost spectrum. The lower barriers will allow developers to use more resources in developing techniques to access the spectrum instead of investing in spectrum.
41
Chapter 2
2.3.3 Dynamic Frequency Selection
OSA represents the general case for accessing the spectrum using “listen before
talk” etiquette. A subset of OSA is DFS, which also is called dynamic channel
selection. DFS is used to prevent a device from accessing a specific band if it is in
use and to prevent co-channel interference of the primary user from a secondary
user. It differs from OSA in that it is not seeking spectrum access—it is determining whether access should be allowed. DFS has been adopted in the 802.11h standard. When the ITU allowed unlicensed WLAN in 2003, 802.11h was developed
in order to satisfy the primary users within the 5.25–5.725 GHz band. DFS detects
other devices using the same radio channel, and it switches WLAN operation to
another channel if necessary. DFS is responsible for avoiding interference with
other devices, such as radar systems and other WLAN devices. Specifically, for
the 5.25–5.725 GHz service, the DFS system must be able to detect a primary user
above a detection threshold of 62 dBm.
2.4 Policy Challenges for Cognitive Radios
The capacity to sense, learn, and adapt to the radio environment provides new
opportunities for spectrum users. However, the same sensing and adaptation also
creates challenges for policy-makers. The primary concern is with the potential to
have nondeterministic behaviors. Nondeterministic behaviors can be created by a
variety of conditions:
●
●
●
The allowance of self-learning mechanisms will create a condition in which the
response to a set of inputs will be changing and thus unknown.
The allowance of software changes will create conditions either from errors
within the software or from rogue software, which can cause the device to not
conform to the transmission rules.
The allowance of frequency and waveform agility will create conditions in
which devices that conform to transmission rules may cause interference due
to mismatch between out-of-band receivers and the in-band transmitter
waveforms.
In addition to nondeterministic behaviors, another primary concern is the
impact of horizontal versus vertical service structure. Vertically integrated services, such as cellular telephony, clearly delineate responsibility for spectrum management to the service provider. The service provider has the sole responsibility
42
Communications Policy and Spectrum Management
for problems, interference, and all other technical and service issues. However,
or horizontally integrated service, which includes device-centric systems that may
be the initial focus for cognitive radio technology, there isn’t a single point of
responsibility for interference and other problems. One example has been the
issue with secondary spectrum markets. The formal responsibility of a device
creating interference is the primary licensee. The rules had to be modified to
allow that responsibility to follow the usage to the secondary licensee when
appropriate. The extrapolation of this approach is problematic when applied
to cognitive radios, as each device is, in essence, a licensee. This is a serious
problem for the policy-makers that can be addressed by rules, technology, or
a combination of both.
This section addresses these concerns with the deployment of cognitive radios.
In particular, the context of cognitive radio for dynamic spectrum access (DSA)
and security concerns are described. The highlighted questions are areas for
research and development (R&D).
2.4.1 Dynamic Spectrum Access
The development of the OSA and DFS technology is straightforward in terms of
radio technology. The discussion of policy impact is much more complex. In both
cases, the dynamic system must interface seamlessly, without interference, to
existing spectrum licensees. Two basic questions need to be addressed:
●
●
Question 1: What are the rights of the license holder to prevent unauthorized
use by an opportunistic device? To answer this question, a clear definition of
interference is needed. The interference definition must include some aspect of
duration and persistence of the interferer. One draconian response could be that
it is better to use less than 1 percent of the spectrum to ensure that a single
packet is never dropped due to interference.
Question 2: What kind of assurance can be provided that the interference will
be for only a finite and precise duration? Answering this question is complicated by the fact that as these systems become more flexible and agile, the
combination of possible configurations grows exponentially.
The ability to move the operations of a radio in frequency is quite enticing.
The operational rules, however, are different in each of the separate frequency
bands. Therefore, a single device must: (1) know of all of the transmission
rules in the bands in which it can operate; (2) have the ability to adjust its
43
Chapter 2
transmission parameters accordingly; and (3) ensure that device will adhere to the
transmission rules.
Defining the Rules for Dynamic Spectrum Access
The policy goal of managing interference is generally defined as to the parameters
for the transmitter (e.g., center frequency, bandwidth, power spectral density,
antenna gain). Interference management is determining the threshold for
producing harmful interference. DSA systems can move in frequency, so
fixed parameters are insufficient. The metric for interference must be defined
to ensure that the DSA device does not cause harmful interference. The primary
questions are:
1. What is an appropriate interference metric?
With an explicit interference metric, the transmission rules could allow operation under a much broader set of transmitter parameters with the explicit goal
of preventing interference above a particular threshold within a radius of the
transmitter. The interference metric could also be contextually sensitive with
respect to location, time, or condition of the RF environment. As the FCC has
said, “Quantitative standards reflecting real-time spectrum use would provide
users with more certainty and, at the same time, would facilitate enforcement.”4
The interference metric must be indicative of the impact of the transmitter
on the surrounding region. In a static and well-defined geometry, a measurement at the transmitter can provide sufficient information from which to
extrapolate the value of the interference metric across a wide area. However,
the highest density of use of the RF spectrum is in areas with complex geometry
and a large number of users, most of which are mobile. Although the technology for propagation models, inclusive of complex environments, is improving
and can provide qualitative results for extrapolation, it is insufficient for quantitative interference analysis. Thus, another important question arises:
2. What kind of in situ measurement system is needed to provide the basis for
interference analysis?
Multiple measurements distributed across the operational area are needed
to accurately measure the interference metric. One possible mechanism to
obtain multiple measurements would be the development of monitoring
4
See [3] Section VI: Interference Avoidance found at http://www.fcc.gov/sptf/reports.html
44
Communications Policy and Spectrum Management
3.
4.
5.
6.
stations such as pollution monitoring devices. Many policy and technical questions arise:
Who would appropriate the funds to develop and deploy such a system? If it is
a federally funded system, should that system be managed by a government or
independent regulatory entity?
Who would ensure the accuracy of the measurements? Who could challenge
the accuracy of the devices?
What are the mechanisms to disseminate the results of the measurements?
Would there be an interface to all service providers and users who could obtain
the data on a near real-time basis? Would the data be broadcast to all devices or
should it be on a request basis?
Liability when the policy is not followed must also be determined. Public safety
communications’ use of interruptible spectrum is one example in which liability
would need to be explicitly addressed. The need for assured communications,
free from avoidable interference, is a paramount requirement for public safety
communications. So the mechanisms for obtaining and releasing spectrum
need to be highly predictable. Beaconing is one such mechanism; that is, either
a service would send a beacon indicating it is using the spectrum or send a
beacon when the band is available for use by another entity. The question then
arises as to how beaconing should be used. Propagation challenges may create
shadow zones that would prevent a nonpublic safety user from hearing the
beacon. So the basic technical challenge is to create the proper signaling technique. Thus, the following question arises:
What policy will assure the public that the system will work, determine who is
liable if it does not work, and fix what type of compensation may be allowed
between the public safety user and the primary licensee?
Safeguards and Incentives for Incumbent Users
Incumbent license holders that are currently using all of their rights to access the
spectrum would initially see no value to having DSA [4]. The desired impact of
DSA is to improve spectrum utilization and thus provide more competition and
products to the market. The challenge is to determine how such impacts can be
obtained while providing both assurances to current license holders and incentives
for increased utilization through DSA.
7. How can assurances be made to current license holders while developing
incentives for DSA?
45
Chapter 2
Technology provides a potential solution to this particular challenge. The
recent development of policy-enabled devices (see Section 2.5.3) can provide
assurance that the device can have “fail-safe” mechanisms to prevent operation
in unauthorized bands.
The problem set that has been put forward essentially requires the policymakers to provide the infrastructure to enable a technology. Generally, this is a
very difficult task because regulatory agencies are not focused on providing an
infrastructure or service to the commercial world. Technologists must actively
address these concerns.
2.4.2 Security
The challenges of employing cognitive radios for the policy community also
include that of ensuring secure device operations. Security in this context includes
enforcement of DSA rules. Enforcement for static systems is already a challenge
due to the amount of resources necessary to authorize equipment, the requirement
of obtaining proof that violations have occurred, and the determination of the violators’ identities. As the systems become more dynamic, there is an increase in the
number of potential interactions that can lead to a violation. Additionally, this
leads to a decrease of the time and spatial scales of these interactions. Both of
these changes will amplify the enforcement challenges.
Equipment Authorization
Initial equipment authorizations have two components that will increase significantly
in complexity with the onset of dynamic policies: evaluation criteria and security
certification. The capacities to modify waveforms, change operating conditions,
and change transmission frequencies all contribute to an exponential growth in
potentially adverse interactions between systems. Exhaustive testing becomes
unrealizable due to the sheer number of combinations. Formal method techniques
have been a focus to solve this problem. Formal methods provide provable atoms
of code that mitigate the requirement for exhaustive testing. However, the maturity
of the technology is inadequate to address the complexity of software embodied in a
cognitive radio. Therefore, the challenge will be to answer a new set of questions:
1. How can realizable test plans be developed for certification that provide sufficient certainty as to the impact of the system on existing, certified equipment?
2. How can a device be certified if the software and hardware come from different
manufacturers?
46
Communications Policy and Spectrum Management
If self-learning mechanisms are to be employed, equipment authorization
becomes even more problematic. Software and hardware certification will not provide sufficient assurances that the device conforms to the operational envelopes.
New mechanisms will need to be developed, such as policy-enabled devices (see
Section 2.5.3).
Software Certification
Software certification and the security of the software are also challenging areas.
Cognitive radio algorithms are written in software that provides the control of
dynamic systems. Software can be modified to allow policies to change on either
a periodic or an aperiodic basis. The security of that software is critical to ensure
that rogue behavior is not programmed into the device. If the consumer can access
the device’s software, then the consumer can instruct the radio to perform outside
of the permitted operational parameters. Thus, basic issues arise such as:
3. How is software protected to ensure that this abhorrent behavior does not
occur?
4. Security is not an absolute but a level of acceptable risk. How much protection
is necessary? How is it tested to ensure that it is sufficiently secure?
5. Who is liable if there is a security failure?
Monitoring Mechanisms
The number of combinations of interactions is high and the mobility and the
agility of future systems is great, so the basic issue remains:
6. How should enforcement systems be developed to observe all of these new
cognitive capabilities?
Three possible mechanisms are suggested: authority-based, network-based,
and infrastructure-based systems. The greatest challenges for development of such
monitoring mechanisms are the equipment and analysis costs and the civil liberty
concerns.
●
5
The authority-based system is for a regulatory agency5 to deploy a national
monitoring system. A secondary advantage of such a national system would be
Or an entity authorized by a regulatory agency.
47
Chapter 2
additional uses. Such a system would have capability of measuring and
reporting spectrum usage, interference environment, and propagation
characteristics.
●
The network-based system is to use the plethora of user devices already
present to monitor the activity of the RF spectrum. The challenge would be
in the methods to:
– provide sufficient confidence in the accuracy of the measurements;
– obtain sufficient geolocation information to make the information
valuable;
– collect and disseminate the information to enforcement organizations within
the regulatory community;
– do so with minimal overhead on network resources.
●
The infrastructure-based system is a combination of authority-based and
network-based systems. The goal would be to use preexisting infrastructure
such as cell towers, Federal Aviation Agency (FAA) towers, and the like. The
challenge would be to obtain the authority to equip each site and to have
priority access to the network from each site.
Again, as with the policy challenges for DSA systems, the challenges for
cognitive radios in the security arena are also quite great. The solution basis is
also a combination of technical and policy initiatives. The technical initiatives
outlined above need to be conducted and vetted through the policy communities.
2.4.3 Communications Policy before Cognitive Radio
This section addresses the role of the communications regulatory bodies on the
development and deployment of cognitive radios as well as the impact of cognitive radios on regulation. Communication regulation is separate and distinct
within every nation. The FCC is the regulatory agency for all wireless communications systems within the United States, except those used by the federal
government (e.g., DoD, National Aeronautics and Space Administration (NASA),
and FAA). Those systems are regulated by National Telecommunications and
Information Administration (NTIA) within the Department of Commerce (DoC).
The primary goal of the regulatory agencies is to promote technology for the public interest while preventing interference between the multitudes of systems
already deployed.
48
Communications Policy and Spectrum Management
2.4.4 Cognitive Radio Impact on Communications Policy
The challenges that face current telecommunication regulators are great.
The rapid development of new techniques to access the spectrum and
applications is quickly outpacing the ability for the policy-maker to react.
This is apparent by how rapidly society is changing its views on spectrum use
and management. The consumer has an expectation of untethered connectivity
with more devices and applications. Technology is also lowering the barriers
for new commercial entrants, therefore increasing both the amount and diversity of the uses for the RF spectrum. In a seemingly contradictory statement,
given the consumer demands for using the RF spectrum, technology is also
challenging the long-held views that spectrum is a scarce resource. Policymakers need a clear, nonreactive vision for future management of the
RF spectrum.
The mirror image is also true. The challenges that face current technologists
and entrepreneurs are also great. The lifespan of an individual market has shortened from decades to years and possibly months. The regulatory approval cycle
can be longer than the market lifespan, creating risks of missing a market altogether. Although the technologists can create new capabilities and products, the
regulators can have a marked impact on their ability to bring those products to
the marketplace. Additionally, technology has become sophisticated and the application market is intertwined with that technology. Regulators must now be technology savvy in order to comprehend all aspects of their decisions. Technologists
and entrepreneurs must now be cognizant of the role regulators have on what
they can develop.
2.4.5 US Telecommunications Policy, Beginning with the Titanic
It is critical to understand the underpinnings of current spectrum management.
The United States provides a good example of the historical context of the development of telecommunications policy. This section briefly reviews the chronology
of events that have shaped US policy.
Frequently, the impetus for new policies is a catastrophe that leads to the
perception of a policy failure. The sinking of the Titanic was the pivotal point
in initiating the development of wireless communications policy. In 1910, the
US Congress had mandated that passenger ships carry wireless telegraphs.
Investigations into the 1911 Titanic disaster indicated that amateur radio operators
caused interference after initial reports of the disaster. This interference, it was
thought, hampered rescue efforts.
49
Chapter 2
Federal Communications Commission
The Radio Act of 1927 established the Federal Radio Commission and set forth
as its intent to “maintain the control of the United States over all the channels
of interstate and foreign radio transmission; and to provide for the use of such
channels, but not the ownership thereof.” The 1927 Act provided that the new
Commission shall, “as public convenience, interest, or necessity requires”
classify radio stations, prescribe the nature of the service, assign bands of
frequencies or wavelengths and determine the power, time, and location of
stations and regulate the kind of apparatus to be used. Licenses were to be
granted by the Commission for a limited duration (3 years for broadcast
licenses and 5 years for all others), but all federal government stations were
to be assigned by the president.
Seven years later, the Communications Act of 1934 abolished the Federal
Radio Commission and transferred the authority for spectrum management to the
newly created FCC. The 1934 Act brought together the regulation of telephone,
telegraph, and radio services within a single independent federal agency. The
1927 Radio Act was absorbed largely intact into Title III of the 1934 Act.
From 1934 to the early 1990s, the US Congress enacted many amendments to
Title III, but there were no fundamental changes to the core provisions that can be
traced back to the 1912 and 1927 acts. However, two noteworthy additions to the
1934 Act inserted in 1983 by Congress were:
●
●
that it is the policy of the United States “to encourage the provision of new
technologies and services to the public” and that anyone who opposes a new
technology or service will have the burden of demonstrating that the proposal is
inconsistent with the public interest; and
“notwithstanding any licensing requirement established in this Act, [the FCC
may] by rule authorize the operation of radio stations without individual
licenses [in certain services].”
Therefore, new technologies will be promoted over existing services,
and license by rule, or what is generally called unlicensed devices, are
permissible.
National Telecommunications and Information Administration
The US DoC also has an important function in telecommunications regulation.
Primarily through the NTIA, the DoC serves as the president’s expert advisor on
50
Communications Policy and Spectrum Management
telecommunications matters and policy. NTIA is charged with reviewing policy
options on behalf of the Executive Branch and communicating proposed policy
decisions to the Congress. NTIA also manages and administers the portion of the
RF spectrum that has been set aside for exclusive use by the federal government.
NTIA is also responsible for coordinating the federal government’s participation
in the ITU WRCs and related national and international meetings.
State Department
The US State Department is the primary representative on foreign policy matters.
Through its Economics Bureau Office of International Communications and
Information Policy, the State Department represents the United States in international telecommunications forums, including bilateral and multilateral negotiations, and before international organizations, such as the ITU. The WRC
delegation is led by the WRC ambassador, whose role is to help negotiate a unified US position at the conference. The US president typically confers the personal rank of ambassador in connection with this special mission for a period not
exceeding 6 months.
2.4.6 US Telecommunications Policy: Keeping Pace with Technology
The regulatory methods have gone through several changes throughout the years.
During the initial era of amplitude modulation (AM) radio, the numbers of systems were small, and interference avoidance was accomplished through frequency
and spatial separation. Although the frequency range was limited to low-frequency
(LF) through high-frequency (HF) bands, it was sufficient to accommodate the
hundreds of stations. These systems included stationary transmitters and mobile
receivers. Characterizing a typical receiver and using propagation models easily
allowed for the computation of potential interference. These calculations worked
very well because the transmitters did not move. The benefit to the system developers was that most of the regulatory requirements were placed upon the transmitters, which were few in comparison to the millions of receivers. As long as the
transmitter complied with the in-band and out-of-band emission limits, interference between systems was prevented. Out-of-band emissions and antenna placement were specifically controlled to assure that the spectral and spatial distances
between two potential interferers were sufficient to prevent interference.
The development of TV and FM radio saw the numbers of systems grow,
as did the number of frequency assignments needed for these systems. If still
limited to the HF bands, then a spectrum crisis would have precipitated. But
51
Chapter 2
fortunately, development of new RF component technology allowed the
operational frequency bands to also grow into the very high frequency (VHF)
and eventually the ultra-high-frequency (UHF) bands. Because the same topology
was still being used—a single transmitter with many receivers—the same type of
computations could be made to ensure a relatively interference-free operation.
Then, in the 1970s and 1980s, transistor and integrated circuit technology
reduced the cost of transmitters. This provided cost-effective transmitters for the
mass market in the form of Citizens’ Band (CB) radios, unlicensed devices, cellular mobile radios, and numerous other devices. Protocols were developed for the
devices and the users in order to provide orderly sharing of the RF channels. The
channels were not all the same bandwidth. Some were wideband to accommodate
high information data rate systems. Most were quite narrow. However, the main
complexity of transmitter mobility had to be addressed. Therefore, the calculations for potential interference were no longer static. Out of this new complexity
came ideas for minimum operational distances between potential interferers.
These placed bounds on the interference. Devices that might be close in spatial
proximity had to rely on spectral separation, and devices close in spectral proximity had to rely on spatial separation.
Although technology continued to march forward, the radio portion of
the US communication policy has remained mostly unchanged for 70 years.
Therefore, current policy is based on the technology that existed for broadcasting: one transmitter to many receivers. The evolution and revolution of radio
technology over the past 15 years has significantly challenged that basis.
Although the exact values can be argued, the United States had an average
of more than two receivers for every person in 1980. The advent of cellular
telephony, data networking, and two-way paging has now added one transceiver
or more for each and every person. But the trends indicate that it will not stop
there. It can be argued that a model similar to what happened to computing is
taking hold in telecommunications. The evolution of technology also incurred an
evolution of computer use from a corporate site, to an office group, to an individual desk, to every home, to every individual, and now to embedded-computing in
many consumer household products. The largest growth occurred when computers
changed from devices that required interaction with the primary user to devices
that work in the background, such as a refrigerator, a TV, a phone, and the automobile. The same trend is occurring within the wireless communications environment. There are now multiple wireless devices per person within the first
adopters, as more consumer products are being developed that use primarily
unlicensed devices.
52
Communications Policy and Spectrum Management
2.5 Telecommunications Policy and Technology Impact on Regulation
Regulations based on static broadcast geometries cannot address the spatial,
numeric, and spectral dynamics of future radio technology. Technologists must
begin to address not only how to construct such new technologies, but also to
address how to bring dynamics into the regulatory framework.
2.5.1 Basic Geometries
Four basic geometries affect the type of technical and social/economic issues that
are addressed in wireless communications policy: fixed or mobile transmitters
combined with fixed or mobile receivers:
●
●
●
●
Fixed Transmitter, Mobile Receiver(s)
Fixed Transmitter, Fixed Receiver(s)
Mobile Transmitter, Fixed Receiver(s)
Mobile Transmitter, Mobile Receiver(s).
Fixed Transmitter, Mobile Receiver(s)
Fixed transmitter, mobile receiver systems include broadcasting, radio position
determination, and standard time and frequency signal services. Broadcasting
comprises a large fraction of the consumer devices such as radio (AM, FM, TV,
etc.). Radio position determination includes radio navigation and radio beaconing
services such as GPS. Standard time and frequency signal services include
WWVB, the National Institute of Standards and Technology (NIST) long-wave
standard time signal, which continuously broadcasts time and frequency signals at
60 kHz. The carrier frequency provides a stable frequency reference traceable to
the national standard. A time code is synchronized with the 60 kHz carrier and is
broadcast continuously at a rate of 1 bit per second (bps). Emission-only devices,
such as those for ISM purposes, including microwave ovens, magnetic resonance
equipment, and industrial heaters, are included in this category. The important
feature of these systems is that there are small numbers of high power transmitters
at fixed and potentially known locations. The economic challenge is to put most
of the complexity (i.e., cost) in the transmitter since the ratio of receivers to transmitters is more than 1 million to 1.
The policy challenge with fixed transmitter, mobile receiver systems is to
determine the allowable transmission parameters (power, location) that prevent
interference at the receivers and provide for the potential of an economically
53
Chapter 2
viable business. The trade-offs for broadcasting services include the number of
stations in a given region and the viability of the service (number of customers).
The trades are exceptionally complex. The station density could be increased by
reducing transmission power and greater frequency reuse. However, that would
decrease the coverage area and thus the number of potential customers. The station density could be increased by using closer band spacing between stations.
However, that would increase the out-of-band rejection by the receivers.
Fixed Transmitter, Fixed Receiver(s)
Fixed transmitter, fixed receiver systems include point-to-point, point-tomultipoint, and radio astronomy services. Both endpoints are in fixed locations
and could either be a one-way (transmitter to receiver) or a two-way (transceiver
to transceiver) configuration. A point-to-point communication system is defined
as having two fixed transceivers. Private operational-fixed microwave may use an
operational-fixed station, and only for two-way communications related to the
licensee’s commercial, industrial, or safety operations. Point-to-multipoint
includes multipoint distribution systems (MDSs) and multichannel, multipoint
distribution systems (MMDSs) that are generally used for one-way data
broadcasting. Originally, the primary MMDS application was “wireless cable”
to deliver TV programs. Advances in antenna development allowed for two-way
digital subscriber link (DSL) applications to be implemented with MMDS.
Radio astronomy is the scientific study of celestial phenomena through
measurement of the characteristics of radio waves emitted by physical processes
occurring in space. The radio telescopes that are used for astronomical work are
extremely large because the signal strength coming from the distant stellar
objects is low and many of the frequencies that are observed are below 3 GHz.
The challenge is in addressing the location of fixed receiver, radio astronomy
systems that take decades to plan and construct. Originally located in places
away from population centers to minimize the potential for interference from
commercial systems, these systems now find themselves surrounded by
population centers. The policy-makers are essentially facing an issue of
whether to keep “radio-free zones” around the telescopes or to find other means
to provide interference-free operation.
Fixed transmitter, fixed receiver systems are the most straightforward to
determine the transmission parameters to prevent interference with other systems.
However, the complexity occurs with mobile transmitters interfering with these
systems. Because the location of the fixed transceivers is generally unknown to
54
Communications Policy and Spectrum Management
mobile users, mobile transmitters can potentially interfere with a receiver operating close to the noise floor due to out-of-band emissions or lack of out-of-band
rejection by the receiver. However, the policy trades are quite straightforward
because the RF environment and the geometry are fixed.
Mobile Transmitter, Fixed Receiver(s)
Mobile transmitter, fixed receiver systems include monostatic active as well as
passive meteorological and Earth exploration systems. These systems are mobile
(either airborne or space based). In the passive sensing configuration, the operational area is unknown, as it has similar characteristics to radio astronomy. In the
active sensing configuration, the transmission location is unknown but the receiver
is co-located with the transmitter.
Mobile transmitter, fixed receiver systems generally employ extremely
sensitive receivers. Due to the mobility of the transmitter, the geographic region
impacted is fixed in size and moves with the transmitter. The policy challenges are
that the amount of frequency needed for these systems is large but the specified
frequency use is very small. Also, the operational frequencies are specific to the
physical attributes of the chemicals that are to be sensed. The sensitivity of the
receivers also requires all adjacent channel systems to have an extremely low outof-band emission. This is usually accomplished through guard bands. The challenge to the policy-maker is to determine the relative values of consumer services
compared with scientific investigation and Earth exploration. Those values
provide input as to whether to find mechanisms to access the unused spectrum
in one location while the sensor is operating in another location.
Mobile Transmitter, Mobile Receiver(s)
Mobile transmitter, mobile receiver systems include a wide range of mobile services as well as portable unlicensed devices. These systems include radiotelephony
(e.g., cellular, personal communication system (PCS), wireless communication,
and specialized mobile radio (SMR)) and private land mobile radio (PLMR) services. PLMR services are for state and local governments, and for commercial and
nonprofit organizations to use for mobile and ancillary fixed communications to
assure the safety of life and property and to improve productivity and efficiency.
Personal radio services include CB radios, Family Radio Service (FRS), and
remote control. Unlicensed devices, also known as licensed-by-rule, license-free,
or “Part 15” devices as denoted in the FCC rules, are included. The important
feature of these systems is that there are potentially large numbers of moderate
55
Chapter 2
(100 W) and extremely large numbers of low power (1 mW–1 W) transceivers that
are mobile. Both the mixture of powers and the unknown geometries between
receivers and transmitters make it impossible to provide absolute assurance of
interference-free operation without specifying a minimum separation distance.
Mobile transmitter, mobile receiver systems involve by far the most complex
geometries for policy-makers to address. Currently, all computations for interference
assume a minimum separation distance between devices (and it differs from device to
device). The assumption is that these distances represent the space in which the user
has full control and thus can directly impact the presence of interference. As with the
mobile transmitter, fixed receiver systems, the mobility of the transmitters creates the
uncertainty of the geometry between the transmitters and receivers. The mobility also
creates spectral regions in time, space, and frequency that are not used. The policymaking challenge is to maximize the use of the spectrum, encourage the development
of new technologies and services, and to provide certainty (spectrum access, interference, etc.) for the service providers to encourage investment.
2.5.2 Introduction of Dynamic Policies
Currently, the government spectrum management rules in the United States are
dynamic with respect to frequency. That is, the rules for particular spectrum-based
services tend to differ based on where a device is authorized to operate in the RF
spectrum. For example, licensed transmitters operating in the radio broadcasting
bands from 88 to 108 MHz must conform to the FM broadcasting rules of Part 73
[5]. The frequencies of such transmitters can be changed only after lengthy regulatory review to ensure such changes will not potentially cause harmful interference. Cognitive radios can change frequency readily, in seconds or milliseconds.
These devices must incorporate aspects of the governing rules or policies from
each of the different spectral areas in which they might operate. New devices that
incorporate wireless fidelity (WiFi®) with mobile telephones and “roam” between
wide area networks (WANs) and local area networks (LANs) are examples of how
multiple spectrum policies can be merged within a single device. The dynamics
are quite limited, yet possible under government and industry policies. The capacity to adjust spectrum policies dynamically opens the new possibility of dynamic
spectrum policies. These policies can be at the device level, by which operational
envelopes can be downloaded and modified by either the regulatory agency or the
primary license holder. The dynamic policies can also be at the system level,
whereas the network policies that are used to optimize performance can now add
the parameter of spectrum access and the associated policies for using a specific
56
Communications Policy and Spectrum Management
band, at a location, with the particular system load. Three operational dimensions
of spectrum policy avail themselves to dynamics—specifically, time, space, and
interference.
●
●
●
Time: An example of using the time dynamic in spectrum policy was exhibited
in the early days of radio. Particular AM stations would cease transmission
late at night and resume early the next morning. Time-based dynamics can be
extended significantly from this example. One extension is to include scheduled/
expected interactions that are quite predictable. These may include secondary
market transactions, in which a secondary provider accesses the spectrum using
a separate network. These also may include the flexible access of a band by the
primary user for a different application, such as reusing a cell site to provide
data telemetry to/from vending machines. A further extension of this concept,
which would be more opportunistic in character and less predictable, yet reliable
for both primary and secondary users, may include using the spectrum for a
short time or within a very limited area. One example is microtransactions within
the secondary market for such “spot” use. Another could be a noncooperative
use of spectrum that is currently not in use. The opportunistic use would exhibit
quick transactions that could be impractical for human intervention. Automated
schemes will be used similar to those used in financial transactions on the
New York Stock Exchange.
Space: Spatial dynamics are depicted in cases where the location of a device
would determine its operational characteristics. One proposal for spatial
dynamics includes the allowance of higher power transmission of unlicensed
devices in rural environs. Another proposal is the use of unlicensed devices in
bands where the device is sufficiently far away from a UHF TV transmitter.
Location sensing would be necessary for the first proposal. Signal strength
sensing would be necessary for the second proposal. In either case, because the
transmitters are stationary, the location information is static. Therefore, once
the boundaries are determined through calculation or measurement, then these
boundaries could be programmed into a device. However, extending the concept to avoid mobile transmitters creates additional complexities. The distance
to mobile transmitters would be constantly changing, and thus more automated
sensing and interference avoidance techniques would be required.
Interference: In contrast to the spatial and temporal dynamics, interference
dynamics would need to understand not only its environment but also the
impact of its own transmission on the surrounding environment. The capacity
57
Chapter 2
to accurately measure and model the environment would be needed. A significant amount of R&D has occurred over the past decade to improve the fidelity
of simulation and modeling of RF propagation. Companies such as Remcom
have products that are examples of those developments. Additionally, device
technology has significantly reduced the cost of RF sensing while also improving
in fidelity.
There are many specific applications of dynamic spectrum policies. In the case
where dynamic policies overlay current static policies, the choice for the device
designer is whether to provide those new capabilities at additional cost for each
device. An example of this is the case of whether to use licensed spectrum, secondary market spectrum, or unlicensed devices. Licensed spectrum has an assured
quality of spectrum access and interference but is associated with higher spectrum
costs. Each of the other choices has less assurance of quality but at lower spectrum costs.
It is easy to expect that with dynamic policies an explosion of new sensing
devices and cooperative networks can be developed. These will be aimed at
providing cost-effective solutions for both licensed and unlicensed uses. The
incorporation of more processing capacity within licensed and unlicensed devices
will present system developers with a large number of choices to provide new
services with variable quality to the consumer.
2.5.3 Introduction of Policy-Enabled Devices
All radio devices are policy-enabled devices at the base level—the policies
(or constraints) are the physical capabilities of the radio (e.g., power output,
frequency range, modes of operation). Cognitive radios offer the ability to
create dynamic-policy-enabled devices, or in other words, cognitive radios can
realize the dynamic usage of frequency bands on an opportunistic basis by
identifying and using the under-utilized spectrum. Policies that determine when
spectrum is to be considered as an available opportunity and that define the possibilities of using these spectrum opportunities must be specified by the regulators
but implemented into the devices. The ability to encode policy into a device,
thus creating a policy-enabled device, can potentially have a profound impact
at the policy level [6], provided the device’s functionality as a radio remains
trustable [7].
The technology for policy-enabled devices is derived from the development of
the Semantic Web (SeW). The SeW is a machine for creating syllogisms. Thus the
58
Communications Policy and Spectrum Management
promise of SeW is to allow third parties to combine assertions to discover things
that are true but not specified directly. The research has focused on the development
of the Resource Description Framework (RDF) and the Web Ontology Language
(OWL). The RDF is a structured environment for representing information about
resources in the World Wide Web (WWW). It is particularly intended for metadata about web resources, such as the title, author, and modification date of a web
page, and so on. OWL can be used to explicitly represent the meaning of terms in
vocabularies and the ontological relationships between those terms.6 Challenges
remain in the development of policy-enabled devices. The functionality of the
policy-enabled device must be “trustable” in its own operation. There has been a
great deal of work in the mathematics of trustable systems and in demonstrating
the impact of a flawed (e.g., not trustable) system (see, e.g., Mitola [7]).
If the ontology for the regulations for wireless communications could be
developed, then devices would be able to directly instantiate a policy into software
and check that policy: (1) for self-consistency and (2) for consistency with other
policies already in use.
Two profound changes for telecommunication policy-making can result from
this technology: (1) policy-makers can now include policies that are distinct in
both space and context and (2) policy-makers can now look at policies that change
with time. The latter is much more significant. The time dimension allows for policy to expire.
These new characteristics allow for more aggressive policy concepts. Current
policy-makers and regulators are extremely cautious because the rules they
develop will be long lived and affect millions of users. Thus the rules are very
conservative and address not only well-understood interference possibilities,
but also rare and statistically remote interference possibilities. In fact, with the
increased density and dynamic behavior of devices, it may be practically impossible (exponentially complex) to ensure interference-free operation. With policyenabled devices, new policies could be tried for shorter periods of time to
determine the impact of the new policies. This would be equivalent to a test of
the spectrum access system.
The ability to have policies instantiated in devices that can change with
respect to location and context is very desirable. Power limits that change with
6
Note that the eXtensible Markup Language (XML) can also be used to define policies, but
devices require the same parser for correct interpretation. OWL is better for distributing policy
information among different types of cognitive radio devices with possibly different interpretation
engines.
59
Chapter 2
respect to signal density or proximity to high usage areas could enable more costeffective deployments of communications to rural and underserved areas. In the
same way, devices in high usage areas would have to be capable of addressing
interference more robustly.
Policy-enabled devices allow for the implementation of dynamic policies.
Such new capabilities for the radios afford new possibilities for the regulator.
2.5.4 Interference Avoidance
The central role of spectrum management is interference management. The
design and operation of RF equipment, including communications and emitting
noncommunications devices, are predicated upon preventing and/or mitigating
electromagnetic interference. In today’s RF environment, interference generally
limits the usable range of communications signals. Interference protection has
always been a core responsibility of communication regulatory bodies such as the
FCC. Section 303(f) of the US Communications Act of 1934 as Amended directs
the FCC to promulgate regulations it deems necessary to prevent interference
between stations, as the public interest shall require [8]. This is still a critical
aspect of the FCC operation. The FCC’s strategic plan for the years 2003–2008
includes as a spectrum-related objective the “vigorous protection against harmful
interference …” [9].
In 2002, the FCC created the Spectrum Policy Task Force (SPTF) to provide
recommendations for future spectrum policy [8]. The SPTF determined that there
are rising concerns that current interference management paradigms will not adequately meet future spectrum demands. Four basic challenges to spectrum management were outlined:
1. many radio communications services have grown substantially in recent years;
2. consumer demand for RF devices has exploded;
3. the technology of waveform flexibility in radios moved from a relatively small
number of waveforms to widely varying signal architectures and modulation
types for voice, video, data, and interactive services;
4. the use of rapidly advancing technology, such as SDR and cognitive radio, will
continue to change the interference landscape.
For example, due to advances in digital signal processing and antenna technology,
communications systems and devices are becoming more tolerant of interference
through their ability to sense and adapt to the RF environment.
60
Communications Policy and Spectrum Management
Additionally, the increased ability of new technologies to monitor their local
RF environment and operate more dynamically than traditional technologies can
provide new mechanisms for interference management. The predictive models
used by government regulators and network operators can be updated, and perhaps eventually replaced, by techniques that take into account and assess actual,
rather than predicted, interference.
2.5.5 Overarching Impact
Regulation of spectrum will undergo revolutionary changes in the near future,
allowing less restricted, more flexible access to spectrum. Such flexible spectrum
usage requires regulation to realize a more open spectrum. Policies that determine
when spectrum is considered as opportunity and that define the possibilities of
using these spectrum opportunities are to be specified. The advent of policies that
change with time and space will allow greater access to the fallow spectrum as
well as lower infrastructure costs. Lower infrastructure costs are obtained through
the ability to have policies change when conditions are warranted. The example of
rural areas being allowed to use a higher transmit power for unlicensed devices
would reduce the number of access points that are needed to service an area. The
optimal density of access points is related to the number of users within the footprint of the access point. Essentially a provider wants a specific number of users
per access point. Therefore, current rules have a highly suboptimal number of
users per access point in low usage areas and a highly suboptimal number of
access points per user in high usage areas. The current set of rules allows the area
covered by the access point to be reduced, but it does not allow the increase of
power. Dynamic policies associated with the state of the RF environment could
allow a more optimal design, and thus a lower infrastructure cost, to be used.
2.6 Global Policy Interest in Cognitive Radios
Since 2000, the interest in how technology has changed both in RF spectrum
needs by the user community and RF spectrum regulatory methods has been
heightened within the radio community. The push to data services within the context of wireless Internet access has been seen as a fundamental shift from the
wireless services to wireless access. The spectrum needs manifested themselves
with the move toward third-generation (3G) services by wireless service providers
as well as the plethora of unlicensed wireless devices e.g., WiFi, cordless phones)
in consumer electronics. These trends actually followed a significant increase of
wireless communications development and usage within the federal users of the
61
Chapter 2
spectrum, such as the military. The inherent limitation of omnidirectional RF
propagation limited the new uses of the spectrum to below 6 GHz.
The development of cognitive radio technology and application concepts,
especially in DSA, has received significant interest worldwide. The greatest interest has been in the United States from both the DoD and the FCC.
2.6.1 Global Interest
Figure 2.8 provides a list of global regulatory activities with respect to cognitive
radios and DSA. Most of the interest outside of the United States and United
Kingdom has been investigatory. A great deal of activity has been ongoing within
the ITU and the European Telecommunications Standards Institute (ETSI)
in developing definitions, standards, and regulatory regimes for using this new
technology.
ETSI is officially responsible for standardization of Information and
Communication Technologies (ICT) within Europe, including telecommunications. ETSI has 688 members from 55 countries inside and outside Europe,
including manufacturers, network operators, administrations, service providers,
research bodies, and users. ETSI plays a major role in developing a wide range of
standards and other technical documentation as Europe’s contribution to worldwide ICT standardization.
2.6.2 US Reviews of Cognitive Radios for Dynamic Spectrum Access
In the United States, a number of federal institutions involved with spectrum regulation, usages, or oversight undertook a wide range of studies. These studies were
broken into two categories: (1) analysis of the use of needs of the RF spectrum
(Government Accountability Office (GAO), Defense Science Board (DSB),
NTIA, and Toffler Associates; see under the sections Government Accountability
Office through Toffler Associates) and (2) investigations of changes needed in the
regulation, policy, and management of the RF spectrum (FCC, Center for
Strategic and International Studies (CSIS), DoC; see under the sections Federal
Communications Commission through US Department of Commerce).
Government Accountability Office
The US GAO exists to support the US legislature through evaluation and investigations of federal programs and policies. This provides additional information for
the US Congress to make oversight, policy, and funding decisions. The GAO has
conducted a series of studies into wireless communications policy. During the
62
Communications Policy and Spectrum Management
●
●
●
●
●
●
●
●
●
United States
– FCC: Report and Order, ET Docket No. 00-47—Authorization and Use of Software Defined Radios; Vanu
Corporation get First Approved SDR in 2004; Notice of Proposed Rulemaking, ET Docket No. 03-108—
Facilitating Opportunities for Flexible, Efficient, and Reliable Spectrum Use Employing Cognitive Radio
Technologies; Notice Proposed Rulemaking, ET Docket No. 04-186—Unlicensed Operation in the TV
Broadcast Bands and Additional Spectrum for Unlicensed Devices Below 900 MHz and in the 3 GHz Band;
Strategic Plan 2006-11—Encouraging the development of new technologies, such as cognitive radio and
dynamic frequency selection; …
Japan
– Ministry of Internal Affairs and Communications (formally MPHPT): Promoting R&D of technologies for
efficient spectrum inclusive of cognitive radio systems to search for unused spectrum.
New Zealand
– Ministry of Economic Development: 2005 Review of Radio Spectrum Policy in New Zealand—the role of
spectrum band managers may be reduced through the use of SDRs, New Zealand’s physical isolation
provides for an ideal position to be a test bed for new wireless technologies, “cognitive or smart” radios will
eventually have the capacity to locate and utilize any unoccupied spectrum, and none of New Zealand’s
current licensing types is well suited for managing SDRs and cognitive radios.
Australia
– Australian Communications and Media Authority (formally ACA and ABA): Vision 20/20 report on future
communications—realizing the future of ubiquitous communications with wireless expected to have an
increasingly central role with increases spectrum sharing and cognitive radio technologies.
Canada
– Canadian Radio–Television and Telecommunications Commission
– Industry Canada: Spectrum Policy Framework for Canada—Implementation of new technologies and new
spectrum management concepts using recent innovations in wireless technology such as cognitive radio and
software defined radio. The department solicited for comments in determining to what extent these
technologies might increase the use and access of the RF spectrum in the future.
ITU
– GSC: GSC-10 (Radiocommunication Items) in 2005 issued a resolution (GSC-10/6) on Global Radio
Standards Collaboration on Wireless Access Systems to encourage collaboration on measurement
techniques and certification requirements for cognitive capabilities including DFS; GSC-10 also called for
accelerated standards development for SDR and cognitive radio; GSC-10 also resolves to study better ways
to manage interference using adaptive frequency agility, listen before transmit, etc.
– ETSI: Considering the impact on SDR of the Radio & Telecommunications Terminal Equipment (R&TTE)
Directive with respect to electromagnetic compatibility (EMC), radio characteristics, nonconforming software,
security, and integrity issues due to potential failures of the software download process.
– ITU-R WP8A and WP8F: Developing definitions and application areas for SDR and cognitive radio
technology. The draft reports indicate that many of the communications administrators around the globe
have begun to investigate the use of these new technologies.
India
– Telecom Regulatory Authority of India: Consultation paper on Issues Relating to Private Terrestrial TV
Broadcasting Service addresses alternative technologies. Comments received state that rural
communications could use cognitive radio technologies to find, in situ, better bands for foliage penetration.
United Kingdom
– Office of Communications (OFCOM): Technology R&D program, initiated in 2005, studies flexible,
multiprotocol, multiband cognitive radio systems. Near-term investigations into band-sharing technologies.
– Commission for Communications Regulation (COMREG) in Ireland: Issued the first cognitive radio/dynamic
spectrum assess test license.
European Union
– Project Team 8 (PT8) (Postal and Telecommunications Administration, Electronic Communications
Committee Working Group): Tasked with developing a report on the regulatory structure needed to enable
the introduction of new radio technologies. A particular focus will be increased opportunities to share
spectrum.
– EU Commission: Published a Forward-Looking Radio Spectrum Policy for the European Union—Second
Annual Report with key initiatives, including the implementation of flexible spectrum usage through the
development of “smart” or cognitive radios. Such efforts will be funded under the EU Research and
Technology Development (RTD) Framework Programme.
Figure 2.8: 2005 International regulatory activities on cognitive radio.
63
Chapter 2
time from 2000 to 2005, the GAO conducted a series of studies on the uses and
expected needs of the spectrum.
Due to the increasing demands for RF spectrum, the GAO was asked to examine whether future spectrum needs can be met given the current regulatory framework. The ensuing report [10] focused mostly on the spectrum management
structure within the United States and the decision-making impediments caused
by having two regulatory agencies focusing on separate constituents: federal users
versus nonfederal users. The significant increase of shared bands between federal
and nonfederal users requires a consensus between the two groups. However, the
report indicated that SDRs and other advanced technologies can potentially alleviate many of the conflicts by making spectrum more plentiful through more efficient access.
In 2004, the GAO performed a study [11] into how agencies within the US
government that access RF spectrum can use advanced technologies to improve
their “spectrum efficiency.” The study clearly acknowledges that technologies
such as “software-defined cognitive radios can be adapted to operate in virtually
any segment of spectrum and, in the future, may be able to adapt to real-time conditions and make use of under-utilized spectrum in a given location and time.”
The report, however, also indicated that the current allocation system does not
allow these technologies to operate. The study also indicates that the lack of
knowledge of the operating environment is an impediment. The report concludes
that new technologies, such as cognitive radio, exhibit exceptional promise for
efficient spectrum utilization but that there are few regulatory requirements or
incentives to employ these technologies.
Defense Science Board
The DSB is a Federal Advisory Committee for the US DoD staffed by technology
and business experts, who conduct studies pertinent to the DoD. The military is
the primary source of R&D funding for advanced communications due to its
changing requirements and the willingness to deploy expensive communications
assets.
The DSB conducted two spectrum-related studies during the period 2000–
2005 on the current and expected DoD needs for accessing the RF spectrum. The
DSB reported in November 2000 [12] that the management of time, space, and
modulation dimensions of the RF spectrum increases the use of the scarce RF
spectrum resource. This control should share the spectrum, in real time, and thus
be more efficient than fixed allocations. These real-time systems dynamically
64
Communications Policy and Spectrum Management
control the assigned frequencies to assure communications. The report further
finds that dynamic frequency assignment by radios that can sense the spectrum
could present unknown problems in spectrum management. It further questions,
the technical issues concerning which technologies to develop and how to apply
these new capabilities for spectrum sharing and dynamic frequency allocation
and/or assignment.
In July 2003, the DSB addressed dynamic access to mobile networks in a
report on wideband RF modulation [13]. In that report, the DSB viewed software
radios as both a risk and an opportunity. The risk lies in the potential of developing and implementing protocols and waveforms that could not be supported by
existing systems and command and control techniques. There is also an interference danger with radios that can change their configuration without knowledge
of how the new transmission waveforms and frequencies would impact other
deployed systems. Even with these risks, the report has a strong recommendation
that the US DoD “increase and focus investment in flexible and adaptive agile
wideband communications technologies to achieve necessary mission capabilities
in a highly dynamic RF communications environment.” That is, invest in
cognitive radios.
National Telecommunications and Information Administration
Although the FCC is the official US regulatory body for RF spectrum, there are
cases in which it does not have jurisdiction. By law, national security use of the
spectrum is under the jurisdiction of the US president. The use of spectrum for
national security has been applied broadly to include federal agencies such as
the DoD, FAA, and Federal Bureau of Investigation (FBI). The jurisdiction of
the federal use of the spectrum has been delegated to the Assistant Secretary of
Commerce for Communications and Information, who is also the Director of
NTIA. NTIA performs numerous studies to investigate the impact of new technology and the needs of the RF spectrum users within federal agencies.
Toffler Associates
In 2001, a study by Toffler Associates provided recommendations as to the next
steps for spectrum policy in the United States [14]. The study focused primarily
on management aspects of spectrum allocations and thus did not address specific
technology impacts such as those from cognitive radio. However, the recommendations from this independent group do address overall issues such as developing
a long-term strategy and how to conduct spectrum reallocations. A primary focus
65
Chapter 2
of the proposed strategy was to develop a framework that anticipated change and
remained versatile. This includes the predication of the needs for future spectrum
allocations as well as the impact of future technologies such as software defined
and cognitive radios.
Federal Communications Commission
The FCC is the US regulatory body for the RF spectrum as per the US
Communications Act. The FCC performs numerous studies on both RF spectrum
needs and technical issues such as interference. In 2002, the FCC chartered a
forward-looking study to investigate the changes occurring in technology and to
recommend how the FCC should develop and implement spectrum policy; more
specifically, the intent of the study was to “identify and evaluate changes in
spectrum policy that will increase the public benefits derived from the use of
radio spectrum. The creation of the SPTF initiated the first ever comprehensive
and systematic review of spectrum policy at the FCC” [15]. A primary goal was
to move from a reactive spectrum management model to one more in line with
a proactive spectrum policy. The SPTF noted that technological advances,
specifically in software defined and cognitive radio, are enabling both the need
to change spectrum policy and the capacity to look at different paradigms to
implement spectrum policy.
The SPTF concluded that technology improvements have shown that the
capacity is limited by the regulatory means used to access the RF spectrum. That
is, access and not technical efficiency is the limiting factor for using the spectrum.
Software defined and cognitive radio technology enables accessing the spectrum
in multiple dimensions, such as time, frequency, bandwidth, and power.
Traditional spectrum management techniques use models and measurements to
predict the interaction between different radio transmitters and receivers. Therefore
operational parameters such as transmission power, out-of-band emission, and the
size of guard bands are selected to ensure interference-free interoperation of
devices. The technological advances in interference rejection and digital coding will
require changes to these parameters if spectrum management is to keep pace with
technology. Without those changes, the operational parameters will be too conservative and thus limit the capacity to use the spectrum efficiently. Additionally, the ability to monitor the RF environment and dynamically alter radio operation makes the
traditional predictive interaction model of management obsolete.
Because operational values can change with technology as well as be modified
in situ, spectrum management will need to change. The current method of
66
Communications Policy and Spectrum Management
explicitly regulating transmission characteristics to prevent interference will need
to move to a rights and responsibilities model in which interference is explicitly
defined and the operational parameters must be set by the device accessing the
spectrum to be within the interference limits.
The SPTF recommended 39 changes to promote more efficient spectrum policies and more intensive use of the RF spectrum. Recommendations in four broad
areas have a direct impact on cognitive radios:
●
●
●
●
Allow for maximum feasible flexibility of spectrum use by both licensed and
unlicensed users.
Adopt a more quantitative approach to interference management based on the
concept of interference temperature.
Promote access to the spectrum through secondary market policies that encourage access for “opportunistic” devices and, where appropriate, permit easements for spectral overlays and underlays.
Promote spectrum access and flexibility in rural areas such as varying power
levels, leasing mechanisms, and geographic licensing.
The overarching impact that the recommendations provided was that the
FCC move from the single-use model of spectrum use to more of a flexible-use
model. In that model, the FCC would clearly define the rights and responsibilities
for access to the RF spectrum and then let the users optimize and interact and/or
trade. It was recommended that policy-makers address the difficult challenge of
defining the appropriate engineering metrics in which to enable such a policy.
Such metrics (e.g., the interference metric, which is discussed in section
Defining the Rules for DSA), could then be implemented directly within a
cognitive radio.
Center for Strategic and International Studies
The CSIS is a private, bipartisan public policy research organization that provides
insights and possible solutions to current and emerging issues. In 2002–2003,
CSIS organized a commission to address spectrum management for the 21st century. The commission focused on four problems for US spectrum management:
(1) lack of long range plans, (2) lack of mechanisms to resolve disputes between
spectrum management organizations, (3) increased challenges in negotiating international spectrum agreements, and (4) the risk to security and economic growth
due to “lag in the development and use of new technologies” [16]. Although the
67
Chapter 2
commission addressed concerns for US policy, the fourth problem is relevant to
all spectrum policy organizations.
The CSIS commission noted that the development of spectrum-based technologies is increasing rapidly. In general, these technologies are initially developed within the military communications communities and then migrate to the
commercial and consumer markets. Thus, technology is providing new ways to
allow more intensive use of the spectrum and could alleviate the perception of a
spectrum shortage. Cognitive radio technology enables the exploitation of gaps in
transmission frequencies and usage times to allow such an increase of spectrum
use. Additionally, such technology can be used to provide more robust behavior to
interference. The commission was concerned that the license-centric spectrum
management policies cannot accommodate such new capabilities.
The CSIS commission had multiple recommendations for improving spectrum
management. Two recommendations addressed organizational structure through
more oversight at the White House, with new National Security Council and
National Economic Council positions and a spectrum advisory board. In order to
address the impact of the pace of technological innovation, the commission also
recommended the establishment of a research consortium to support the government and private sectors. This consortium would establish goals for research in
spectrum innovations as well as provide a platform for resolving technical disputes that arise as technology changes.
US Department of Commerce
In June 2003, the US president issued a memorandum recognizing how the RF
spectrum contributes to “significant innovation, job creation and economic
growth.” He then created a Spectrum Policy Initiative, chaired by the DoC, to
develop recommendations for improving spectrum management policies and procedures [17, 18]. The subsequent report from the initiative indicates: “Given the
increase in new and innovative radio communication systems seeking access to
the spectrum, the most challenging issue is interference problems inherent in
using the latest technologies.” The report continues to address the challenges from
technology by stating, “The unpredictable nature of … ingenuity is not to be
solved—it is a reality to be embraced …”
Many of the recommendations put forth by the initiative are for modifications
to the structure of the spectrum management community. However, there are two
recommendation areas that could significantly impact the development of cognitive radios: a Spectrum Sharing Innovation Testbed and Spectrum Management
Tools. The report acknowledges the need for more sharing between spectrum
68
Communications Policy and Spectrum Management
users is inevitable due to the increased need for spectrum and the available technology to enable such sharing. In fact, 54 percent of the spectrum allocations
below 3 GHz are shared and 94 percent below 300 GHz are shared. The initiative
report recommends the development of a Spectrum Sharing Innovation Testbed,
which consists of spectrum that will be set aside exclusively to test and evaluate
new methods for spectrum sharing.
The initiative also recognized that there will be a continued explosion of spectrum uses, and that spectrum managers need the proper tools to be effective and
efficient. The employment of cognitive radios was specifically mentioned as a
technology whose promise and limitations should be better understood by both
the FCC and the NTIA.
2.7 Summary
By 2005, the policy community had embraced the utility of cognitive radios from
the vantage points of new applications as well as new policy regimes. Due to the
lack of operational prototypes, the community has been in a preparatory phase in
developing definitions and potential operation envelopes in which cognitive radios
may function. The greatest amount of interest is in the DSA applications because
these are seen as opportunities to provide more services and new technologies
without having to allocate new spectrum.
This chapter has presented a review of relevant technical and policy definitions. Key among them are the following:
●
●
●
●
Cognitive radios consist of four new capabilities: (1) flexibility, to change waveform and configuration; (2) agility, to change the spectral band in which to
operate; (3) sensing, to observe the state of the system; and (4) networking, to
communicate between multiple nodes for aggregating capacity.
The GSC within the ITU is developing a definition of cognitive radios as well
as extensions to also cover policy based and DFS radios.
Currently, there are three basic frequency assignment methods: command and
control, auctions, and protocols and etiquettes. Command and control has been
employed the most, but the use of auctions has been growing in popularity
since its inception in 1994. Protocols and etiquettes are generally applied for
unlicensed or license-free devices.
OSA provides a new mechanism to dynamically obtain frequency assignments
through sensing open spectral regions and adapting frequency selection in a
69
Chapter 2
cognitive radio. This ability has been under development within the technical
community since 1999 and within the regulatory community since 2002.
●
Interruptible spectrum is a special subset of frequency assignment when aperiodic increase in spectral assignment is obtained by interruption of part of the
band of another service.
This chapter has also addressed the policy-relevant issues that need to be
addressed if cognitive radios are to be used to their fullest capacity:
●
●
●
●
Policy challenges for cognitive radios include addressing nondeterminism of
self-learning algorithms, verification/validation of software, and the impact of
waveform flexibility on out-of-band receivers.
DSA policy must address the question of the rights of the license holder to prevent unauthorized use by an opportunistic device.
DSA uses waveform flexibility and frequency agility to optimize performance.
Interference metrics to determine the impact of waveforms on receivers is
needed to provide a quantitative method to limit and/or eliminate interference.
Multiple mechanisms are employed to insure security operation of a cognitive
radio: equipment authorization, software certification, and in situ monitoring of
a transmitter. Equipment authorization and software certification become more
impractical as the number of operations and/or software states grows exponentially. In situ monitoring is a developing technology.
Finally, this chapter provided a short tutorial as to the relevant regulatory roles with
cognitive radios:
●
●
●
●
US spectrum policy agencies are the NTIA and the FCC.
Four basic geometries determine the type of technical and economic issues
that are addressed in communications policy: (1) fixed transmitter, mobile
receiver(s); (2) fixed transmitter, fixed receiver(s); (3) mobile transmitter,
mobile receiver(s); and (4) mobile transmitter, fixed receiver(s).
Spectrum management policies can be selected to optimize system and network
performance given current spectrum activity and interference properties in
space and time.
Significant global interest in cognitive radio technology and policy exists, including Japan, New Zealand, Australia, Canada, the United States, the International
Telecommunications Union, India, the United Kingdom, and the European Union.
70
Communications Policy and Spectrum Management
References
[1] Kolodzy, P, “Dynamic Spectrum Policies: Promises and Challenges,” CommLaw
Conspectus, 2004.
[2] http://www.ntia.doc.gov/osmhome/allochrt.pdf
[3] http://www.med.govt.nz/rsm/img/spectrum-chart.jpg
[4] R. Lynch, “Exploring Technology Frontiers—Is There a Network in Your Future?”
Keynote address, Dyspan Conference, Baltimore, Maryland, November 9, 2005.
[5] 47 C.F.R. § 73.200–73.600, SubPart B FM Broadcast Stations.
[6] L. Berlemann, S. Mangold and B. Walke, “Policy-Based Reasoning for Spectrum
Sharing in Cognitive Radio Networks,” Proceedings of the 1st IEEE International
Symposium on New Frontiers in Dynamic Spectrum Access Networks (DySPAN),
November, 2005.
[7] J. Mitola III, “Software Radio Architecture: A Mathematical Perspective,” IEEE
Journal on Selected Areas in Communications, Vol. 17, No. 4, April 1999.
[8] “FCC Spectrum Policy Task Force Report,” www.fcc.gov/sptf, 2002.
[9] See Federal Communications Commission Strategic Plan FY 2003–FY 2008, available at http://www.fcc.gov/omd/strategicplan2003-2008.pdf
[10] US General Accounting Office, “Comprehensive Review of U.S. Spectrum
Management with Broad Stakeholder Involvement Is Needed,” GAO-03-277,
January, 2003.
[11] US General Accounting Office, “Better Knowledge Needed to Take Advantage of
Technologies That May Improve Spectrum Efficiency,” GAO-04-666, May, 2004.
[12] US Department of Defense, “Report of the Defense Science Board Task Force on
DoD Frequency Spectrum Issues,” Washington, DC, 2000.
[13] US Department of Defense, Report of the “Defense Science Board Task Force on
Wideband Radio Frequency Modulation,” Washington, DC, 2000.
[14] S. Kenney, J. O’Connor and R. Szafranski, “Creating the Future of Spectrum
Allocation”; contact through tofflerassociates@toffler.com; 2001.
[15] http://www.fcc.gov/sptf/reports.html
[16] R. Galvin, J. Schlesinger, “Spectrum Management for the 21st Century,” CSIS
Press, Washington, DC, 2003.
[17] US Department of Commerce, “Spectrum Policy for the 21st Century—The
President’s Spectrum Policy Initiative: Report 1,” Washington, DC, DoC, 2004.
[18] US Department of Commerce, “Spectrum Policy for the 21st Century—The
President’s Spectrum Policy Initiative: Report 2,” Washington, DC, DoC, 2004.
71
This page intentionally left blank
CHAPTER 3
The Software Defined Radio as a
Platform for Cognitive Radio
Pablo Robert
Mobile and Portable Radio Research Group, Bradley Department
of Electrical and Computer Engineering, Virginia Tech,
Blacksburg, VA, USA
Bruce A. Fette
Communications Networks Division, General
Dynamics C4 Systems, Scottsdale, AZ, USA
3.1 Introduction
This chapter explores both the hardware and software domains of software defined
radio (SDR). The span of information covered is necessarily broad; therefore, it
focuses on some aspects of hardware and software that are especially relevant to
SDR design. Beyond their obvious differences, hardware and software analyses
have some subtle differences. In general, hardware is analyzed in terms of its
capabilities. For example, a particular radio frequency (RF) front-end (RFFE) can
transmit up to a certain frequency, a data converter can sample a maximum bandwidth, and a processor can provide a maximum number of million instructions per
second (MIPS). Software, in contrast, is generally treated as an enabler. For example, (1) a signal processing library can support types of modulation, (2) an OS can
support multithreading, or (3) a particular middleware implementation can support
naming structures. Given this general form of viewing hardware and software, this
chapter presents hardware choices as an upper bound on performance, and software as a minimum set of supported features and capabilities.
Cognitive radio (CR) assumes that there is an underlying system hardware and
software infrastructure that is capable of supporting the flexibility demanded by the
cognitive algorithms. In general, it is possible to provide significant flexibility
73
Chapter 3
with a series of tunable hardware components that are under the direct control of
the cognitive software. In the case of a cognitive system that can support a large
number of protocols and air interfaces, it is desirable to have a generic underlying
hardware structure.
The addition of a series of generalized computing structures underlying the
cognitive engine implies that the cognitive engine must contain hardware-specific
knowledge. With this hardware-specific knowledge, the cognitive engine can then
navigate the different optimization strategies that it is programmed to traverse.
The problem with such knowledge is that a change in the underlying hardware
would require a change in the cognitive engine’s knowledge base. This problem
becomes exacerbated when one considers porting the engine to other radio
platforms. For example, there could be a research and development platform that
is used to test a variety of cognitive algorithms. As these algorithms mature, it
is desirable to begin using these algorithms in deployed systems. Ideally, one
would just need to place the cognitive engine in the deployed system’s management
structure. However, if no abstraction were available to isolate the cognitive
engine from the underlying hardware, the cognitive engine would need to be
modified to support the new hardware platform. It is clear that an abstraction is
desirable to isolate the cognitive engine from the underlying hardware. The
abstraction of hardware capabilities for radio software architecture is a primary
design issue.
SDR is more than just an abstraction of the underlying hardware from the
application. SDR is a methodology for the development of applications, or waveforms in SDR parlance, in a consistent and modular fashion such that both software and hardware components can be readily reused from implementation to
implementation. SDR also provides the management structure for the description,
creation, and tear-down of waveforms. In several respects, SDR offers the same
capabilities supported by OSs; SDR is actually a superset of the capabilities
provided by an OS. SDR must support a variety of cores, some of which may be
deployed simultaneously in the same system. This capability is like a distributed
OS designed to run over a heterogenous hardware environment, where heterogenous in this context means not only general purpose processors (GPPs), but also
digital signal processors (DSPs), field-programmable gate arrays (FPGAs), and
custom computing machines (CCMs). Furthermore, SDR must support the RF and
intermediate frequency (IF) hardware that is necessary to interface the computing
hardware with radio signals. This support is largely a tuning structure coupled
with a standardized interface. Finally, SDR is not a generic information technology (IT) solution in the way that database management is. SDR deals explicitly
74
The Software Defined Radio as a Platform for Cognitive Radio
with the radio domain. This means that context is important. This context is most
readily visible in the application programming interface (API), but is also apparent
in the strict timing requirements inherent to radio systems, and the development
and debugging complexities associated with radio design.
This chapter is organized as follows: Section 3.2 introduces the basic radio hardware architecture and the processing engines that will support the cognitive function.
Section 3.3 discusses the software architecture of a SDR. Section 3.4 discusses SDR
software design and development. At present, many SDRs utilize a Software
Communications Architecture (SCA) as a middleware to establish a common framework for waveforms, and the SCA is covered in some detail in this section. Section
3.5 discusses applications as well as the cognitive functionality and languages that
support cognitive software as an application. Section 3.6 discusses the development process for SDR software components. Section 3.7 then discusses cognitive
waveform development. Finally, Section 3.8 presents a summary of the chapter.
3.2 Hardware Architecture
The underlying hardware structure for a system provides the maximum bounds for
performance. The goal of this section is to explore hardware for SDR from a radio
standpoint. Figure 3.1(a) shows a basic radio receiver. As one example based on
the basic radio receiver architecture, Figure 3.1(b) shows a design choice made
possible by digital signal processing techniques, in which the sampling process
Symbol
Recovery
Carrier
Recovery
Filter
Symbol
Detector
Demodulator
FEC
MAC
Processing
Network
Stack
Application
AGC
(a)
Carrier
Recovery
Filter
Symbol
Recovery
Phase
Detector
Symbol
Detector
Demodulator
FEC
MAC
Processing
Network
Stack
Application
AGC
(b)
Figure 3.1: (a) Data flow and component structure of a generalized coherent radio receiver;
(b) Data flow and component structure of a generalized coherent radio receiver designed for
digital systems, with sampling as a discrete step (see Design Choices section on page 81 for
an explanation).
75
Chapter 3
for digital signal processing can be placed in any of several locations and still
provide equivalent performance.
3.2.1 The Block Diagram
The generic architecture tour presented here traces from the antenna through the
radio and up the protocol stack to the application.
RF Externals
Many radios may achieve satisfactory performance with an antenna consisting of
a passive conductor of resonant length, or an array of conductors that yield a beam
pattern. Such antennas range from the simple quarter-wavelength vertical to the
multi-element Yagi and its wide bandwidth cousin, the log periodic antenna.
Antennas used over a wide frequency range will require an antenna tuner to
optimize the voltage standing wave radio (VSWR) and corresponding radiation
efficiency. Each time the transceiver changes frequency, the antenna tuner will
need to be informed of the new frequency. It will either have a prestored table
derived from a calibration process, and then adjust passive components to match
the tuning recommendations of the table, or it will sense the VSWR and adapt the
tuning elements until a minimum VSWR is attained.
Some modern antennas include a number of passive components spread over
the length of the radiating elements that are able to present reasonable VSWR
performance without active tuning. The best such units today span a frequency
range of nearly 10:1. However, for radios that are expected to span 2 MHz–2 GHz,
it is reasonable to expect that the radio will need to be able to control a switch to
select an appropriate antenna for the frequency band in which the transceiver is
currently operating. Where beam antennas are used, it may also be necessary for
the radio to be able to manage beam pointing. In some cases, this is accomplished
by an antenna rotator, or by a dish gimbals. The logic of how to control the
pointing depends greatly on the application. Examples include: (1) exchanging the
global positioning system (GPS) position of each transceiver in the network, as
well as tracking the three-dimensional (3-D) orientation of the platform on which
the antenna is mounted so that the antenna pointing vector to any network
member can be calculated; (2) scanning the antenna full circle to find the maximum
signal strength; (3) dithering the antenna while tracking peak performance;
or (4) using multiple receive feed elements and comparing their relative strength.
Another common antenna is the electronically steered antenna. Control
interfaces to these antennas are quite similar in function; however, due to the
76
The Software Defined Radio as a Platform for Cognitive Radio
ability to rapidly steer electronically, many of these antennas update their steering
angle as rapidly as once every millisecond. Thus, response time of the control
interfaces is critical to the expected function.
The most sophisticated antenna is the multiple input, multiple output (MIMO)
antenna. In these antennas, the interface boundary between the radio and the antenna
is blurred by the wide bandwidth and large number of complex interfaces between
the beam-steering signal processing, the large number of parallel RF front-end
receivers, and the final modem signal processing. For these complex antennas, the
SDR Forum is developing interface recommendations that will be able to anticipate
the wide variety of multi-antenna techniques currently used in modern transceivers.
Another common external component is the RF power amplifier (PA). Typically,
the external PA needs to be told to transmit when the transceiver is in the transmit
mode and to stop transmitting when the transceiver is in the receive mode. A PA will
also need to be able to sense its VSWR, delivered transmit power level, and its temperature, so that the operator can be aware of any abnormal behavior or conditions.
It is also common to have a low-noise amplifier (LNA) in conjunction with an
external PA. The LNA will normally have a tunable filter with it. Therefore, it is
necessary to be able to provide digital interfaces to the external RF components to
provide control of tuning frequency, transmit/receive mode, VSWR and transmit
power-level sensing, and receive gain control.
In all of these cases, a general-purpose SDR architecture must anticipate the
possibility that it may be called upon to provide one of these control strategies for
an externally connected antenna, and so must provide interfaces to control
external RF devices. Experience has shown Ethernet to be the preferable standard
interface, so that remote control devices for RF external adapters, switches, PAs,
LNAs, and tuners can readily be controlled.
RF Front-End
The RF front-end consists of the receiver and the transmitter analog functions. The
receiver and transmitter generally consist of frequency up converters and down
converters, filters, and amplifiers. Sophisticated radios will choose filters and
frequency conversions that minimize spurious signals, images, and interference
within the frequency range over which the radio must work. The front-end design
will also maximize the dynamic range of signals that the receiver can process,
through automatic gain control (AGC). For a tactical defense application radio, it
is common to be able to continuously tune from 2 MHz to 2 GHz, and to support
analog signal bandwidths ranging from 25 kHz to 30 MHz. Commercial applications
need a much smaller tuning range. For example, a cell phone subscriber unit
77
Chapter 3
might tune only 824 MHz–894 MHz and might need only one signal bandwidth.
With a simplified design, the designer can eliminate many filters, frequency
conversions, and IF bandwidth filters, with practical assumptions.
The RF analog front-end amplifies and then converts the radio carrier
frequency of a signal of interest down to a low IF so that the receive signal can be
digitized by an analog-to-digital converter (ADC), and then processed by a DSP
to perform the modem function. Similarly, the transmitter consists of the modem
producing a digital representation of the signal to be transmitted, and then a
digital-to-analog converter (DAC) process produces a baseband or IF representation
of the signal. That signal is then frequency shifted to the intended carrier frequency, amplified to a power level appropriate to close the communication link over
the intended range, and delivered to the antenna. If the radio must transmit and
receive simultaneously, as in full duplex telephony, there will also be some filtering
to keep the high-power transmit signal from interfering with the receiver’s ability
to detect and demodulate the low-power receive signal. This is accomplished by
complex filters usually using bulk acoustic wave or saw filters at frequencies below
2 GHz, or yttrium-iron-garnet (YIG) circulators at frequencies above 2 GHz.
The typical software-defined RFFE begins notionally with receiving a signal
and filtering the signal to reflect the range of frequency covered by the intended
signals. For a spread spectrum wideband code division multiple access
(WCDMA) signal, this could be up to 6 MHz of bandwidth. In order to assure that
the full 6 MHz is presented to the modem without distortion, it is not unusual for
the ADC to digitize 12 MHz or so of signal bandwidth. In order to capture
12 MHz of analog signal bandwidth set by the IF filters without aliasing artifacts,
the ADC will probably sample the signal at rates above 24 million samples per
second (Msps). After sampling the signal, the digital circuits will shift the
frequency of the RF carrier to be centered as closely as possible to 0 Hz direct
current (DC) so that the signal can again be filtered digitally to match the exact
signal bandwidth. Usually this filtering will be done with a cascade of several
finite impulse response (FIR) filters, designed to introduce no phase distortion
over the exact signal bandwidth. If necessary, the signal is despread, and then
refiltered to the information bandwidth, typically with an FIR filter.
Analog-to-Digital Converters
The rate of technology improvement versus time has not been as profound for
A/D converters as for digital logic. The digital receiver industry is always looking
for wider bandwidth and greater dynamic range. Successive approximation ADCs
were replaced by flash converters in the early 1990s, and now are generally
78
The Software Defined Radio as a Platform for Cognitive Radio
replaced with sigma-delta ADCs. Today’s ADC can provide up to 105 Msps at
14-bit resolution. Special-purpose ADCs have been reported to provide sample
rates over 5 giga samples per second (Gsps) at 8-bit resolution.
State-of-the-art research continues to push the boundaries of analog-to-digital
(A/D) performance with a wide variety of clever techniques that shift the
boundaries between DSP and ADC.
Modem
After down-conversion, filtering, and equalization, the symbols are converted to
bits by a symbol detector/demodulator combination, which may include a
matched filter or some other detection mechanism as well as a structure for
mapping symbols to bits. A symbol is selected that most closely matches the
received signal. At this stage, timing recovery is also necessary, but for symbols
rather than the carrier. Then the output from the demodulator is in bits.
The bits that are represented by that symbol are then passed to the forward
error correcting function to correct occasional bit errors. Finally, the received
and error-corrected bits are parsed into the various fields of message, header,
address, traffic, etc. The message fields are then examined by the protocol layers
eventually delivering messages to an application (e.g., Web browser or voice
coder (VoCoder)), thus delivering the function expected by the user.
SDRs must provide a wide variety of computational resources to be able to
gracefully anticipate the wide variety of waveforms, they may ultimately be
expected to demodulate. Today we would summarize that a typical SDR should be
able to provide at least 266 MIPS of GPP, 32 Mbytes of random access memory
(RAM), 100 MIPS of DSP, and 500 K equivalent gates of FPGA-configurable
logic. More performance and resources are required for sophisticated waveforms
or complex networking applications. Typically, the GPP will be called upon to
perform the protocol stack and networking functions, the DSP will perform the
physical layer modulation and demodulation, the FPGA will provide timing and
control as well as any special-purpose hardware accelerators that are particularly
unique to the waveform. It appears that SDR architectures will continue to evolve
as component manufacturers bring forward new components that shift the boundaries of lowest cost, highest performance, and least power dissipation.
Forward Error Correction
In some instances, the demodulated bits are passed on to a forward error correction (FEC) stage for a reduction in the number of bit errors received. One of the
79
Chapter 3
interesting aspects of FEC is that it can be integrated into the demodulation
process, such as in trellis-coded modulation; or it can be closely linked to demodulation, as in soft decoding for convolutional codes; or it can be an integral part of
the next stage, medium access control (MAC) processing.
Medium Access Control
MAC processing generally includes framing information, with its associated
frame synchronization structures, MAC addressing, error detection, link management structures, and payload encapsulation with possible fragmentation/defragmentation structures. From this stage, the output is bits, which are input to the
network-processing layer. The network layer is designed for end-to-end connectivity support. The output of the network layer is passed to the application layer,
which performs some sort of user functions and interface (speaker/microphone,
graphical user interface, keypad, or some other sort of human-computer interface).
User Application
The user’s application may range from voice telephony, to data networking, to text
messaging, to graphic display, to live video. Each application has its own unique
set of requirements, which, in turn, translate into different implications on the performance requirements of the SDR.
For voice telephony today, the dominant mode is to code the voice to a moderate data rate. Data rates from 4800 bps to 13,000 bps are popular in that they
provide excellent voice quality and low distortion to the untrained listener. The
digital modem, in turn, is generally more robust to degraded link conditions than
analog voice would be under identical link conditions.
Another criterion for voice communications is low latency. Much real experience with voice communications makes it clear that if the one-way delay for each
link exceeds 50 milliseconds,1 then users have difficulty in that they expect a
response from the far speaker and, hearing none, they begin to talk just as the
response arrives, creating frequent speech collisions. In radio networks involving
ad hoc networking, due to the delay introduced by each node as it receives and
retransmits the voice signaling, it can be quite difficult to achieve uniformly low
delay. Since the ad hoc network introduces jitter in packet delivery, the receiver
must add a jitter buffer to accommodate a practical upper bound in latency of late
1
Some systems specify a recommended maximum latency limit, such as 150 milliseconds for
ITU-T G114.
80
The Software Defined Radio as a Platform for Cognitive Radio
packets. All of this conspires to add considerable voice latency. In response, voice
networks have established packet protocols that allocate traffic time slots to the
voice channels, in order to guarantee stable and minimal latency. In much the
same way, video has both low error rate and fixed latency channel requirements,
and thus networking protocols have been established to manage the quality
requirements of video over wireless networks. Many wireless video applications
are designed to accept the bit errors but maintain the fixed latency.
In contrast, for data applications, the requirement is that the data must
arrive with very few or absolutely no bit errors; however, latency is tolerated
in the application.
Voice coding applications are typically implemented on a DSP. The common
voice coding algorithms require between 20 and 60 MIPS and about 32 Kbytes of
RAM. Voice coding can also be successfully implemented on GPPs, and will typically require more than six times the instructions per second (100–600 MIPS) in
order to perform both the multiply—accumulate signal processing arithmetic and
the address operand fetch calculations.
Transmitting video is nearly 100 times more demanding than voice, and is
rarely implemented in GPP or DSP. Rather, video encoding is usually implemented on special purpose processors due to the extensive cross-correlation
required to calculate the motion vectors of the video image objects. Motion
vectors substantially reduce the number of bits required to faithfully encode the
images. In turn, a flexible architecture for implementing these special purpose
engines is the use of FPGAs to implement the cross-correlation motion-detection
engines.
Web browsing places a different type of restriction on an SDR. The typical
browser needs to be able to store the images associated with each Web page in
order to make the browsing process more efficient, by eliminating the redundant
transmission of pages recently seen. This implies some large data cache function,
normally implemented by a local hard drive disk. Recently, such memories are
implemented by high-speed flash memory as a substitute for rotating electromechanical components.
Design Choices
Several aspects of the receiver shown in Figure 3.1(a) are of interest. One of the
salient features is that the effect of all processing between the LNA and the FEC
stage can be largely modeled linearly. This means that the signal processing chain
does not have to be implemented in the way shown in Figure 3.1(a). The carrier
recover loop does not have to be implemented at the mixer stage. It can just as
81
Chapter 3
easily be implemented immediately before demodulation. Another point to note
is that no sampling is shown in Figure 3.1(a). It is mathematically possible to
place the sampling process anywhere between the LNA and the FEC, giving
the designer significant flexibility. An example of such design choice selection
is shown in Figure 3.1(b), where the sampling process is shown as a discrete
step.
The differences seen between Figures 3.1(a) and 3.1(b) are not at a functional
level; they are implementation decisions that are likely to lead to a dramatically
different hardware structure for equivalent systems. The following discussion on
hardware concentrates on processing hardware selections because a discussion of
RF and IF design considerations are beyond the scope of this book.
Several key concepts must be taken into consideration for the front-end of the
system, and it is worthwhile to briefly mention them here. From a design standpoint, signals other than the signal of interest can inject more noise in the system
than was originally planned, and the effective noise floor may be significantly
larger than the noise floor due to the front-end amplifier.
One example of unpredicted noise injection is the ADC conversion process,
which can inject noise into the signal through a variety of sources. The ADC
quantization process injects noise into a signal. This effect becomes especially
noticeable when a strong signal that is not the signal of interest is present in an
adjacent channel. Even though it will be removed by digital filtering, the stronger
signal sets the dynamic range of the receiver by affecting the AGC, in an effort to
keep the ADC from being driven into saturation. Thus, the effective signal-tointerference and noise ratio (SINR) of the received signal is lower than might
otherwise be expected from the receiver front-end. To overcome this problem and
others like it, software solutions will not suffice, and flexible front-ends that are
able to reject signals before sampling become necessary. Tunable RF components,
such as tunable filters and amplifiers, are becoming available, and the SDR design
that does not take full advantage of these flexible front-ends will handicap the
final system performance.
3.2.2 Baseband Processor Engines
The dividing line between baseband processing and other types of processing,
such as network stack processing, is arbitrary, but it can be constrained to be
between the sampling process and the application. The application can be
included in this portion of processing, such as a VoCoder for a voice system or
image processing in a video system. In such instances, the level of signal processing
82
The Software Defined Radio as a Platform for Cognitive Radio
is such that it may be suitable for specialized signal processing hardware, especially in demanding applications such as video processing.
Four basic classes of programmable processors are available today: GPPs,
DSPs, FPGAs, and CCMs.
General Purpose Processors
GPPs are the target processors that probably come to mind first to anyone writing
a computer program. GPPs are the processors that power desktop computers and
are at the center of the computer revolution that began in the 1970s. The landscape of microprocessor design is dotted with a large number of devices from a
variety of manufacturers. These different processors, while unique in their own
right, do share some similarities, namely, a generic instruction set, an instruction
sequencer, and a memory management unit (MMU).
There are two general types of instruction sets: (1) machines with fairly broad
instruction sets, known as complex instruction set computers (CISCs); and (2)
machines with a narrow instruction set, known as reduced instruction set computers (RISCs). Generally, the CISC instructions give the assembly programmer
powerful instructions that address efficient implementation of certain common
software functions. RISC instruction sets, while narrower, are designed to produce
efficient code from compilers. The differences between the CISC and RISC are
arbitrary, and both styles of processors are converging into a single type of
instruction set. Regardless of whether the machine is CISC or RISC, they both
share a generic nature to their instructions. These include instructions that perform
multiplication, addition, or storage, but these instruction sets are not tailored to a
particular type of application. In the context of CR, the application in which we
are most interested is signal processing.
The other key aspect of the GPP is the use of a MMU. Because GPPs are
designed for generic applications, they are usually coupled with an OS. This OS
creates a level of abstraction over the hardware, allowing the development of
applications with little or no knowledge of the underlying hardware. Management
of memory is a tedious and error-prone process, and in a system running multiple
applications, memory management includes paging memory, distributed programming and data storage throughout different blocks of memory. An MMU allows
the developer to “see” a contiguous set of memory, even though the underlying
memory structure may be fragmented or too difficult to control in some other
fashion (especially in a multitasking system that has been running continuously
for an extended period of time). Given the generic nature of the applications that
83
Chapter 3
run on a GPP, an MMU is critical because it allows the easy blending of different
applications with no special care needed on the developer’s part.
Digital Signal Processors
DSPs are specialized processors that have become a staple of modern signal processing systems. In large part, DSPs are similar to GPPs. They can be programmed with a high-level language such as C or C and they can run an OS.
The key difference between DSPs and GPPs comes in the instruction set and
memory management. The instruction set of a DSP is customized to particular
applications.
For example, a common signal processing function is a filter, an example of
which is the Direct Form II infinite impulse response (IIR) filter. Such a filter is
seen in Figure 3.2.
Figure 3.2: Structure and data flow for Direct
Form II IIR filter as a basis for an estimate on
computational load.
b0
x [n]
y [n]
a1
x 1
b1
a2
z 1
b2
an
z 1
bn
As seen in Figure 3.2, the signal path is composed of delays on the incoming
signal of one sample, z1, and the multiplication of each delayed sample by a
coefficient of the polynomial describing either the poles (a) or zeros (b) of the filter. If each delayed sample is considered to be a different memory location, then
to quickly implement this filter, it is desirable to perform a sample shift in the circular buffer, perform a multiply and an add that multiplies each delayed sample
times the corresponding polynomial coefficients, and store that result either in the
output register, in this case y[n], or into the register that is added to the input, x[n].
The algorithm in Figure 3.2 has several characteristics. Assuming that the filter is of length N (N-order polynomials), then the total computation cost for this
84
The Software Defined Radio as a Platform for Cognitive Radio
algorithm can be computed. To optimize radio modem performance, filters are frequently designed to be FIR filters with only the b (all zeros) polynomial. In order
to implement extremely efficient DSP architectures, most DSP chips support performing many operations in parallel to make an FIR filter nearly a single instruction that is executed in a 1 clock cycle instruction loop. First, there is a loop
control mechanism. This loop control has a counter that has to be initialized and
then incremented in each operation, providing a set of N1 operations. Within the
loop, there is also an evaluation of the loop control state; this evaluation is performed N times. Within the loop, a series of memory fetch operations have to be
performed. In this case, there are N1 accesses to get the coefficient and store the
result. There are additional N1 circular accesses of the sample data plus the
fetch operation of the new data point, yielding a total of 2N2 memory accesses.
Finally, there are the arithmetic operations: the coefficient multiplied by the signal
sample, and the accumulation operation (including the initial zero operation)
resulting in 2N1 operations. Therefore, the DSP typically performs on the order
of 6N operations per instruction clock cycle.
Assuming that a GPP performs a memory fetch, memory store, index update,
comparison, multiplication, or addition in each computer clock cycle, then the
GPP would require 6N3 clock cycles per signal sample. Using these assumptions, then an FIR filter with 32 filter taps and a signal sampled at 100 ksps would
yield (6 32 3) 100 103 19.5 MIPS. Therefore, just for the filter mentioned, almost 20 MIPS are required for the GPP to successfully filter this narrowband signal. In reality, a GPP is likely to expend more than one cycle for each of
these operations. For example, an Intel Pentium 4 floating point multiply occupies
6 clock cycles, so the given performance is a lower bound on the GPP MIPS load.
A DSP, in contrast, has the ability to reduce some of the cycles necessary to
perform the given operations. For example, some DSPs have single-cycle MAC
(multiple and accumulate). The Motorola 56001 is an example DSP that performs
single instruction multiply and accumulate with zero overhead looping. These
reductions result in a computational total of 3N3. Given these reductions, the
computation load for the DSP is now (3 32 3) 100 103 9.9 MIPS.
Given its customized instruction set, a DSP can implement specialized signal
processing functionality with a significantly fewer clock cycles than the more
generic GPP processors. However, GPPs are attempting to erode this difference
using multiple parallel execution arithmetic logic units so that they can perform
effective address calculations in parallel with arithmetic operations. These are
called superscalar architectural extensions. They are also attempting to raise the
performance of GPP multipliers through use of many pipeline stages, so that the
85
Chapter 3
multiplication steps can be clocked at higher speed. This technique is called a
highly pipelined architecture.
Field-Programmable Gate Arrays
FPGAs are programmable devices that are different in nature from GPPs and
DSPs. An FPGA comprises some discrete set of units, sometimes referred to as
logical elements (LE), logic modules (LM), slices, or some other reference to a
self-contained Boolean logical operation. Each of these logical devices has at least
one logic unit; this logic unit could be one or more multipliers and one or more
accumulators, or a combination of such units in some FPGA chip selections.
Logical devices are also likely to contain some memory, usually a few bits. The
developer then has some freedom to configure each of these logical devices,
where the level of reconfigurability is arbitrarily determined by the FPGA manufacturer. The logical devices are set in a logic fabric, a reconfigurable connection
matrix that allows the developer to describe connections between different logical
devices. The logic fabric usually also has access to some additional memory that
logical devices can share. Timing of the different parts of the FPGA can be controlled by establishing clock domains. This allows the implementation of multirate
systems on a single FPGA.
To program an FPGA, the developer describes the connections between logical devices as well as the configuration of each of these logical devices. The final
design that the developer generates is closer to a circuit than a program in the traditional sense, even though the FPGA is ostensibly a firmware programmable
device. Development for the FPGA is done by using languages such as very highspeed integrated circuit (VHSIC) Hardware Design Language (VHDL), which can
also be used to describe application-specific integrated circuits (ASICs), essentially non-programmable chips. Variants of C exist, such as System-C, that allow
the developer to use C-like constructs to develop FPGA code, but the resulting
program still describes a logic circuit running on the FPGA.
The most appealing aspect of FPGAs is their computational power. For example, the Virtex II Pro FPGA by Xilinx has a total of 6460 slices, where each slice
is composed of two look-up tables, two flip-flops, some math logic, and some
memory. To be able to implement 802.11a, a communications standard that is
beyond the abilities of any traditional DSP in 2005, would require approximately
3000 slices, or less than 50 percent of the FPGA’s capabilities, showing the importance of a high degree of parallelism in the use of many multiply accumulators to
implement many of the complex-waveform signal processes in parallel.
86
The Software Defined Radio as a Platform for Cognitive Radio
From a performance standpoint, the most significant drawback of an FPGA is
that it consumes a significant amount of power, making it impractical for battery
powered handheld subscriber solutions. For example, the Virtex II Pro FPGA
mentioned above is rated at 2648 mW of power expenditure, whereas a low-power
DSP such as the TMS320C55x is rated at 65–160 mW, depending on clock speed
and version, and its high-performance cousin, the TMS320C64x, is rated at
250–1650 mW, depending on clock speed and version.
3.2.3 Baseband Processing Deployment
One of the problems that the CR developer will encounter when designing a system is attempting to determine what hardware should be included in the design
and what baseband processing algorithm to deploy into which processors. The initial selection of hardware will impose limits on maximum waveform capabilities,
while the processing deployment algorithm will present significant run-time challenges if left entirely as an optimization process for the cognitive algorithm. If the
processing deployment algorithm is not designed correctly, it may lead to a suboptimal solution that, while capable of supporting the user’s required quality of
service (QoS), may not be power efficient, quickly draining the system batteries
or creating a heat dissipation problem.
The key problem in establishing a deployment methodology is one of scope.
Once a set of devices and algorithm performance for each of these devices has
been established, there is a finite set of possibilities that can be optimized. The
issue with such optimization is not the optimization algorithm itself; several algorithms exist today for optimizing for specific values, such as the minimum mean
square error (MMSE), maximum likelihood estimation (MLE), genetic algorithms, neural nets, or any of a large set of algorithms. Instead, the issue is in
determining the sample set over which to perform the optimization.
There is no openly available methodology for establishing the set of possible
combinations over which optimization can occur. However, at least one approach
has been suggested by Neel et al. [1] that may lead to significant improvements in
deployment optimization. The proposed methodology is partitioned into platformspecific and waveform-specific analyses. The platform-specific analysis is further
partitioned into two types, DSP/GPP, and FPGA. The platform-specific analysis is
as follows:
1. Create an operations audit of the target algorithms (number and type of
operations).
87
Chapter 3
2. For DSP:
(a) Create a set of target devices.
(b) Establish cycle-saving capabilities of each target device.
3. For FPGA:
(a) Create a set of devices.
(b) Establish mapping between different FPGA families. This mapping can be
done on the basis of logical devices, available multiplies per element, or
another appropriate metric.
(c) Find logical device count for each target algorithm. FPGA manufacturers
usually maintain thorough libraries with benchmarks.
(d) Use mapping between devices to find approximate target load on devices
when a benchmark is not available.
Once the platform-specific analysis is complete, the developer now has the
tools necessary to map specific algorithms onto a set of devices. Note that at this
stage, there is no information as to the suitability of each of these algorithms to
different platforms, since a base clock rate (or data rate) is required to glean that
type of information.
Given the platform-specific information assembled above, it is now possible to
create performance estimates for the different waveforms that the platform is
intended to support:
1. Create a block-based breakdown of the target waveform (using target algorithms from step 1 as building blocks).
2. Breakdown target waveform into clock domains.
3. Estimate time necessary to complete each algorithm.
(a) In the case of packet-based systems, this value is fairly straightforward.2
(b) In the case of stream-based systems, this value is the allowable latency.
4. Compute number of operations per second (OPS) needed for each algorithm.
2
If the received signal is blocked into a block of many signal samples, and then the receiver
operates on that block of signal samples through all of the receive signal processes, the process
can be imagined to be a signal packet passing through a sequence of transforms. In contrast,
if each new signal sample is applied to the entire receive process that is referred to here as a
stream-based process.
88
The Software Defined Radio as a Platform for Cognitive Radio
5. Create a set of devices from the platform-specific phase that meet area and cost
parameters (or whatever other parameters are appropriate for a first cut). This
set of devices can be very large. At this stage, the goal is to create a set of
devices or combination of devices that meet some broad criteria.
6. Cycle through the device set.
(a) Attempt to map algorithms onto given devices in set.
(i) For DSP:
(1) Make sure that OPS calculated in step 4 of the waveform-specific
analysis are reduced by cycle-saving capabilities outlined in step
2b of the platform-specific analysis.
(2) The result of the algorithm map is a MIPS count for each device.
(ii) For FPGA:
(1) Mapping of the algorithm is a question of the number of occupied
logical devices.
(2) Make sure that clock domains needed for algorithms can be
supported by the FPGA.
(b) If a solution set of MIPS and/or LE exists for the combination of devices
in the set, then save the resulting solution set for this device set; if a solution does not exist, discard the device set.
7. Apply appropriate optimization algorithm over resulting solution set/device set
from step 6. Additional optimization algorithms include power budgets, and
performance metrics.
The process described yields a coarse solution set for the hardware necessary
to support a particular set of baseband processing solutions. From this coarse
set, traditional tools can be used to establish a more accurate match of resources
to functionality. Furthermore, the traditional tools can be used to find a solution
set that is based on optimization criteria that are more specific to the given needs.
3.2.4 Multicore Systems and System-on-Chip
Even though several computing technologies have some promise over the horizon,
such as quantum computing, it is an undeniable fact that silicon-based computing
such as multicore systems and system-on-chip (SoC) will continue to be the bedrock
of computing technology. Unfortunately, as technology reaches transistors under
100 nm, the key problems become the inability to continue the incremental pace
of clock acceleration as well as significant problems in power dissipation. Even
89
Chapter 3
though the number of gates per unit area has roughly doubled every 18 months
since the 1970s, the amount of power consumed per unit area has remained
unchanged. Furthermore, the clocks driving processors have reached a plateau,
so increases in clock speed have slowed significantly.
In order to overcome the technology problems in fabrication, a design shift
has begun in the semiconductor industry. Processors are moving away from
single-core solutions to multicore solutions, in which a chip is composed of more
than one processing core. Several advantages are evident from such solutions.
First, even though the chip area is increasing, it is now populated by multiple
processors that can run at lower clock speeds. In order to understand the ramifications of such change, it is first important to recall the power consumption of an
active circuit as:
PCfV2
(3.1)
As shown in Eq. (3.1), the power dissipated, P, by an active circuit is the product of the switching activity, , the capacitance of the circuit, C, the clock speed,
f, and the operating voltage, V, squared. It is then clear from the above equation
that the reduced clock speed results in a proportional reduction in power consumption. Furthermore, since a lower operating frequency means that a lower
voltage is needed to operate the device, the reduction in the operating voltage produces a reduction in power consumption that follows a quadratic curve.
One of the principal bottlenecks in processor design is the input/output interface from data inside the chip to circuits outside the chip. This interface tends to
be significantly slower than the data buses inside the chip. By reducing this interface capacitance and voltage swing for intercore communications, system efficiency grows. Furthermore, communication through shared memory is now
possible within the chip. This capability can greatly increase the efficiency of the
design.
3.3 Software Architecture
Software is subject to differences in structure that are similar to those differences
seen in the hardware domain. Software designed to support baseband signal processing generally does not follow the same philosophy or architecture that is used
for developing application-level software. Underlying these differences is the need
to accomplish a variety of quite different goals. This section outlines some of the
key development concepts and tools that are used in modern SDR design.
90
The Software Defined Radio as a Platform for Cognitive Radio
3.3.1 Design Philosophies and Patterns
Software design has been largely formalized into a variety of design philosophies,
such as object-oriented programming (OOP), component-based programming
(CBP), or aspect-oriented programming (AOP). Beyond these differences is the
specific way in which the different pieces of the applications are assembled,
which is generally referred to as a design pattern. This section describes design
philosophies first, providing a rationale for the development of different
approaches. From these philosophies, the one commonly used in SDR will be
expanded into different design patterns, showing a subset of approaches that are
possible for SDR design.
Design Philosophies
Four basic design philosophies are used for programming today: linear programming (LP), OOP, CBP, and AOP.
Linear Programming
LP is a methodology in which the developer follows a linear thought process for
the development of the code. The process follows a logical flow, so this type of
programming is dominated by conditional flow control (such as “if-then” constructs) and loops. Compartmentalized functionality is maintained in functions,
where execution of a function involves swapping out the stack, essentially changing the context of operation, performing the function’s work, and returning results
to the calling function, which requires an additional stack swap. An analogy of LP
is creating a big box for all items on your desktop, such as the phone, keyboard,
mouse, screen, headphone, can of soda, and picture of your attractive spouse, with
no separation between these items. Accessing any one item’s functionality, such
as drinking a sip of soda, requires a process to identify the soda can, isolate the
soda can from the other interfering items, remove it from the box, sip it, and then
place it back into the box and put the other items back where they were. C is the most
popular LP language today, with assembly development reserved for a few brave
souls who require truly high speed without the overhead incurred by a compiler.
Object-Oriented Programming
OOP is a striking shift from LP. Whereas LP has data structures—essentially variables that contain an arbitrary composition of native types such as float or integer—
OOP extends the data structure concept to describe a whole object. An object is a
91
Chapter 3
collection of member variables (such as in a data structure) and functions that can
operate on those member variables. From a terminology standpoint, a class is an
object’s type, and an object is a specific instance of a particular class. There are
several rules governing the semantics of classes, but they generally allow the
developer to create arbitrary levels of openness (or visibility), different scopes,
different contexts, and different implementations for function calls that have the
same name. OOP has several complex dimensions; additional information can be
found elsewhere (e.g., Budd [2] and Weisfeld [3]).
The differences inherent in OOP have dramatic implications for the development of software. Extending the analogy from the previous example, it is now
possible to break up every item on your desktop into a separate object. Each
object has some properties, such as the temperature of your soda, and each object
also has some functions that you can access to perform a task on that object, such
as drinking some of your soda. There are several languages today that are OOP
languages. The two most popular ones are Java and C, although several other
languages today are also OOP languages.
Component-Based Programming
CBP is a subtle extension of the OOP concept. In CBP, the concept of an object is
constrained; instead of allowing any arbitrary structure for the object, under CBP
the basic unit is now a component. This component comprises one or more
classes, and is completely defined by its interfaces and its functionality. Again
extending the previous example, the contents on the desktop can now be organized into components. A component could be a computer, where the computer
component is defined as the collection of the keyboard, mouse, display, and the
actual computer case. This particular computer component has two input interfaces, the keyboard and the mouse, and one output interface, the display. In future
generations of this component, there could be additional interfaces, such as a set
of headphones as an output interface, but the component’s legacy interfaces are
not affected by this new capability. Using CBP, the nature of the computer is irrelevant to the user as long as the interfaces and functionality remain the same. It is
now possible to change individual objects within the component, such as the keyboard, or the whole component altogether, but the user is still able to use the computer component the same as always. The primary goal of CBP is to create
stand-alone components that can be easily interchanged between implementations. Note that CBP is a coding style, and there are no mainstream languages that
are designed explicitly for CBP.
92
The Software Defined Radio as a Platform for Cognitive Radio
Even though CBP relies on a well-defined set of interfaces and functionality,
these aspects are insufficient to guarantee that the code is reusable or portable
from platform to platform. The problem arises not from the concept, but from the
implementation of the code. To see the problem, it is important now to consider
writing the code describing the different aspects of the desktop components that
we described before, in this case a computer. Conceptually, we have a component
that contains an instance of a display, keyboard, mouse, headphone, and computer.
If one were to write software emulating each of these items, not only would the
interfaces and actual functional specifications need to be written, but also a wide
variety of housekeeping functions, including, for example, notification of failure.
If any one piece of the component fails, it needs to inform the other pieces that it
failed, and the other pieces need to take appropriate action to prevent further malfunctions. Such a notification is an inherent part of the whole component, and
implementing changes in the messaging structure for this notification on any one
piece requires the update of all other pieces that are informed of changes in state.
These types of somewhat hidden relationships create a significant problem for
code reuse and portability because relationships that are sometimes complex need
to be verified every time that code is changed. AOP was designed to attempt to
resolve this problem.
Aspect-Oriented Programming
AOP allows for the creation of relationships between different classes. These relationships are arbitrary, but can be used to encapsulate the housekeeping code that
is needed to create compatibility between two classes. In the messaging example,
this class can include all the messaging information needed for updating the state
of the system. Given this encapsulation, a class such as the headphone in the
ongoing example can be used not only in the computer example, but also in other
systems, such as a personal music player, an interface to an airplane sound system, or any other appropriate type of system. The relationship class encompasses
an aspect of the class; thus, context can be provided through the use of aspects.
Unlike CBP, AOP requires the creation of new language constructs that can associate an aspect to a particular class; to this end, there are several languages
(AspectJ, AspectC , and Aspect#, among others).
Design Philosophy and SDR
The dominant philosophy in SDR design is CBP because it closely mimics the
structure of a radio system, namely the use of separate components for the different
93
Chapter 3
functional blocks of a radio system, such as link control or the network stack.
SDR is a relatively new discipline with few open implementation examples, so as
the code base increases and issues in radio design with code portability and code
reuse become more apparent, other design philosophies may be found to make
SDR software development more efficient.
Design Patterns
Design patterns are programming methodologies that a developer uses within the
bounds of the language the developer happens to be using. In general, patterns
provide two principal benefits: they help in code reuse and they create a common
terminology. This common terminology is of importance when working on teams
because it simplifies communications between team members. As will be shown
in the next section, some architectures, such as the SCA, use patterns. For example, the SCA uses the factory pattern for the creation of applications. In the context of this discussion, patterns for the development of waveforms and the
deployment of cognitive engines will be shown. The reader is encouraged to
explore formal patterns using available sources (e.g., Gamma et al. [4], Shalloway
and Trott [5], or Kerievsky [6]).
3.4 SDR Development and Design
From the discussion above, it is clear that a software structure following a collection of patterns is needed for efficient large-scale development. In the case of
SDR, the underlying philosophy coupled with a collection of patterns is called an
architecture or operating environment. There are two open SDR architectures,
GNURadio and SCA.
3.4.1 GNURadio
GNURadio [7] is a Python-based architecture (see section Python) that is
designed to run on general-purpose computers running the Linux OS. GNURadio
is a collection of signal processing components and supports primarily one RF
interface, the universal software radio peripheral (USRP), a four channel up- and
down-converter board coupled with ADC and DAC capabilities. This board also
allows the use of daughter RF boards. GNURadio in general is a good starting
point for entry-level SDR and should prove successful in the market, especially in
the amateur radio and hobbyist market. GNURadio does suffer from some limitations—namely, (1) that it is reliant on GPP for baseband processing, thus limiting
94
The Software Defined Radio as a Platform for Cognitive Radio
its signal processing capabilities on any one processor, and (2) it lacks distributed
computing support, limiting solutions to single-processor systems, and hence limiting its ability to support high-bandwidth protocols.
3.4.2 Software Communications Architecture
The other open architecture is the SCA, sponsored by the Joint Program Office
(JPO) of the US Department of Defense (DoD) under the Joint Tactical Radio
System (JTRS) program. The SCA is a relatively complex architecture that is
designed to provide support for secure signal processing applications running on
heterogenous, distributed hardware. Furthermore, several solutions are available
today that provide support for systems using this architecture, some of which are
available openly, such as Virginia Tech’s OSSIE [8] or Communications Research
Center’s SCARI [9], providing the developer with a broad and growing support base.
The SCA is a component management architecture; it provides the infrastructure to create, install, manage, and de-install waveforms as well as the ability to
control and manage hardware and interact with external services through a set of
consistent interfaces and structures. There are some clear limits to what the SCA
provides. For example, the SCA does not provide such real-time support as maximum latency guarantees or process and thread management. Furthermore, the
SCA also does not specify how particular components must be implemented, what
hardware should support what type of functionality, or any other deployment strategy that the user or developer may follow. The SCA provides a basic set of rules
for the management of software on a system, leaving many of the design decisions
up to the developer. Such an approach provides a greater likelihood that the developer will be able to address the system’s inherent needs.
The SCA is based on some underlying technology to be able to fulfill two
basic goals, namely, code portability and reuse. In order to maintain a consistent
interface, the SCA uses Common Object Request Broker Architecture (CORBA)
as part of its middleware. CORBA is software that allows a developer to perform
remote procedure calls (RPCs) on objects as if they resided in the local memory
space even if they reside in some remote computer. The specifics of CORBA are
beyond the scope of this book, but a comprehensive guide on the use of CORBA
is available [10].
In recent years, implementations of CORBA have appeared on DSP and
FPGA, but traditionally CORBA has been written for GPP. Furthermore, system
calls are performed through an OS, requiring an OS on the implementation. In the
case of the SCA, the OS of choice is a portable operating system interface
95
Chapter 3
(POSIX) PSE-52 compliant OS, but CORBA need not limit itself to such an OS.
Its GPP-centric focus leads to a flow for the SCA, as seen in Figure 3.3.
Black
Transmission
Security
Figure 3.3: CORBA-centric structure for the SCA. The DoD implementation of the SCA
requires that the RF modem (black) side of an SDR be isolated from the side where the plain
text voice or data is being processed (red), and that this isolation be provided by the
cryptographic device.
As seen in Figure 3.3, at the center of the SCA implementation is an OS,
implying but not requiring the use of a GPP. Different pieces of the SCA are
linked to this structure through CORBA and the interface definition language
(IDL). IDL is the language used by CORBA to describe component interfaces,
and is an integral part of CORBA. The different pieces of the system are attached
together by using IDL. An aspect of the SCA that is not obvious from Figure 3.3
is that there can be more than one processor at the core, since CORBA provides
location independence to the implementation. Beyond this architecture constraint
are the actual pieces that make up the functioning SCA system, namely SCA and
legacy software, non-CORBA processing hardware, security, management software, and an integrated file system.
96
The Software Defined Radio as a Platform for Cognitive Radio
The SCA is partitioned into four parts: the framework, the profiles, the API,
and the waveforms. The framework is further partitioned into three parts, base
components, framework control, and services. Figure 3.4 is a diagram of the different classes and their corresponding parts.
Port
PropertySet
PortSupplier
LifeCycle
TestableObject
ResourceFactory
Resource
Base Components
ApplicationFactory
Application
0...*
0...*
Waveform Control
Device
0...*
DomainManager
1...*
AggregateDevice
DeviceManager
LoadbleDevice
System Control
Framework
1
ExecutableDevice
0...1
Control
FileManager
FileSystem
File
Hardware Control
Framework Services
Figure 3.4: Classes making up the SCA core framework.
The SCA follows a component-based design, so the whole infrastructure
revolves around the goal of creating, installing, managing, and de-installing the
components making up a particular waveform.
Base Components
The base class of any SCA-compatible component is the Resource class. The
Resource class is used as the parent class for any one component. Since components are described in terms of interfaces and functionality, not the individual
makeup of the component, it follows that any one component comprises one or
more classes that inherit from the Resource. For example, the developer may create a general Filter category for components. In this example, the filter may be
implemented as an FIR or IIR filter. These filters can then be further partitioned
into specific Filter implementations up to the developer’s discretion, but for the
purposes of this example, assume that the developer chooses to partition the filters
97
Chapter 3
only into finite or infinite response times (rather than some other category such as
structure (e.g., Butterworth, Chebyshev) or spectral response (e.g., low-pass, highpass, notch). If the developer chooses to partition the Filter implementation into
FIR and IIR, then a possible description of the Filter family of classes can be that
shown in Figure 3.5.
Figure 3.5: Sample
filter component class
hierarchy.
Resource
Filter
IIR
FIR
The Resource base class has only two member methods, start() and stop(), and
one member attribute, identifier. A component must do more than just start and
stop, it must also be possible to set the component’s operational parameters, initialize the component into the framework and release the component from the
framework, do some sort of diagnostic, and connect this component to other
components. In order to provide this functionality, Resource inherits from the
following classes: PropertySet, LifeCycle, TestableObject, and PortSupplier, as
seen in Figure 3.6.
Figure 3.6: Resource class
parent structure. The resource
inherits a specialized
framework functionality.
PropertySet
LifeCycle
TestableObject
PortSupplier
Resource
PropertySet
The PropertySet class provides a set of interfaces that allow other classes to both
configure and query the values associated with this Resource. As part of the creation process of the component, the framework reads a configuration file and sets
the different values associated with the component through the PropertySet interface. Later on, during the run-time behavior of the component, the Resource’s
properties can be queried through the query() interface and reconfigured with the
configure() interface.
98
The Software Defined Radio as a Platform for Cognitive Radio
LifeCycle
The LifeCycle parent class provides the Resource with the ability to both initialize
and release the component from the framework through the initialize() and
releaseObject() interfaces. Initialization in this context is not the same as configuring the different values associated with the component. In this context, initialization sets the component to a known state. For example, in the case of a filter, the
initialization call may allocate the memory necessary to perform the underlying
convolution and it may populate the filter polynomial from the component’s set of
properties. The releaseObject performs the complimentary function to initialize.
In the case of a filter, it would deallocate the memory needed by the component.
One aspect of releaseObject that is common to all components is that it unbinds
the component from CORBA. In short, releaseObject performs all the work necessary to prepare the object for destruction.
TestableObject
Component test and verification can take various forms, so it is beyond the scope
of the SCA to outline all tests that are possible with a component. However, the
SCA does provide a simple test interface through the use of the TestableObject
parent class. TestableObject contains a single interface, runTest(). runTest() takes
as an input parameter a number signifying the test to be run and a property structure to provide some testing parameters, allowing the inclusion of multiple tests in
the component. Even though the interface provides the ability to support multiple
tests, it is fundamentally a black box test structure.
PortSupplier
The final capability that parent classes provide to a Resource is the ability to connect to other components. Connections between components are performed
through Ports (discussed below), not through the Resource itself. The use of
PortSupplier allows the Resource to return one of any number of Ports that are
defined within the component. To provide this functionality, the only interface
provided by PortSupplier is getPort(). As its name implies, getPort returns the port
specified in the method’s arguments.
ResourceFactory
The SCA provides significant latitude concerning how an instance of a Resource
can be created. In its broadest context, a Resource can be created at any time by
99
Chapter 3
any entity before it is needed by the framework for a specific waveform. This
wide latitude is not necessarily useful because sometimes the framework needs
the Resource to be explicitly created. In the cases in which it needs to be explicitly created, the specifications provide the framework with the ResourceFactory
class. The ResourceFactory class has only three methods, createResource(),
releaseResource(), and shutdown(). The createResource function creates an
instance of the desired Resource and returns the reference to the caller, and the
releaseResource function calls the releaseObject interface in the specified Resource.
The shutdown function terminates the ResourceFactory. How the ResourceFactory
goes about actually creating the Resource is not described in the specifications.
Instead, the specifications provide the developer with the latitude necessary to
create the Resource in whichever way seems best for that particular Resource.
Port
The Port class is the entry point and, if desired, the exit point of any component.
As such, the only function calls that are explicitly listed in the Port definition are
connectPort() and disconnectPort(). All other implementation-specific member
functions are the actual connections, which, in the most general sense, are guided by
the waveform’s API. A component can have as many Ports as it requires. The implementation of the Port and the structure that is used to transfer information between
the Resource (or the Resource’s child) to the Port and back is not described in the
specifications and is left up to the developer’s discretion. Section 3.3.1 describes
the development process and describes a few patterns that can be used to create
specific components.
Framework Control
The base component classes are the basic building blocks of the waveform, which
are relatively simple; the complexity in component development arrives in the
implementation of the actual component functionality. When developing or
acquiring a framework, the bulk of the complexity is in the framework control
classes. These classes are involved in the management of system hardware, systemwide and hardware-specific storage, and deployed waveforms. From a high
level, the framework control classes provide all the functionality that one would
expect from an OS other than thread and process scheduling.
From Figure 3.4, the framework control classes can be partitioned into three
basic pieces: hardware control, waveform control, and system control. The following sections describe each of these pieces in more detail.
100
The Software Defined Radio as a Platform for Cognitive Radio
Hardware Control
As discussed in Section 3.1, a radio generally comprises a variety of different
pieces of hardware. As such, the framework must include the infrastructure necessary to support the variety of hardware that may be used. From a software placement point of view, not all hardware associated with a radio can run software.
Thus, the classes that are described to handle hardware can run on the target hardware itself, or it can run in a proxy fashion, as seen in Figure 3.7. When used as a
proxy, the hardware control software allows associated specialized hardware to
function as if it were any other type of hardware.
Figure 3.7: Specialized hardware controlled
by programmable interface through a proxy
structure.
Programmable HW
OS
SCA
Framework
SCA
Controller
Driver
Specialized
HW
The needs of the hardware controller in the SCA are similar to the needs of
components supporting a waveform. Furthermore, the component model closely
fits the concept of hardware. Thus, the device controllers all inherit from the
Resource base class. There are four device controllers: Device, LoadableDevice,
ExecutableDevice, and AggregateDevice. Each of these classes is intended to support hardware with increasingly complex needs.
The most fundamental hardware is the hardware that performs some hardwired functionality and that may or may not be configurable. Such hardware has
some inherent capacities that may be allocated for specific use. An example of
such a piece of hardware is an ADC. The ADC is not programmable, but it is conceivable for an ADC to be tunable, meaning the developer may set the sampling
101
Chapter 3
rate or number of quantization bits. In such a case, the available capacity of this
tunable hardware depends on whether it is being used for data acquisition of a
particular signal or over a particular band. To this end, the only two functions that
the Device class includes are allocateCapacity and deallocateCapacity.
Beyond the simple ability to allocate and deallocate capacities, a particular
piece of hardware may have the ability to load and unload binary images.
These images are not necessarily executable code, but they are images
that configure a particular piece of hardware. For example, an FPGA, once
loaded with a bit image, is configured as a circuit. The FPGA is never
“run,” it just acts as a circuit. The LoadableDevice class was created to deal
with such hardware, where LoadableDevice inherits from Device. As would
be expected, the only two functions that the LoadableDevice class contains are
load() and unload().
Finally, there is the more complex hardware that not only has capacities that can
be allocated and deallocated as well as memory that can be loaded and unloaded
with binary images, but it also can execute a program from the loaded binary image.
Such hardware is a GPP or a DSP. For these types of processors, the SCA uses the
ExecutableDevice class. Much like an FPGA, a GPP or DSP has capacities that can
be allocated or deallocated (like ports), and memory that can hold binary images,
so the ExecutableDevice class inherits from the LoadableDevice class. As would
be expected, the two functions that ExecutableDevice supports are execute() and
terminate().
DeviceManager
The different hardware controllers behave as stand-alone components, so they
need to be created and controlled by another entity. In the case of the SCA, this
controller is the DeviceManager. The DeviceManager is the hardware booter; its
job is to install all the appropriate hardware for a particular box and to maintain
the file structure for that particular set of hardware. The boot-up sequence for the
DeviceManager is fairly simple, as shown in Figure 3.8.
As seen in Figure 3.8, when an instance of the DeviceManager is created, it
installs all the Devices described in the DeviceManager’s appropriate profile and
installs whatever file system is appropriate. After installing the hardware and file
system, it finds the central controller, in this case the DomainManager, and installs
all the devices and file system(s).
In a distributed system with multiple racks of equipment, the DeviceManager
can be considered to be the system booter for each separate rack, or each separate
board. As a rule, a different DeviceManager is used for each machine that has a
102
The Software Defined Radio as a Platform for Cognitive Radio
Repeat for
All HW
Platforms
Instantiate
DeviceManager
and Read
Configuration
File
Power-up
Install Devices
and File System
Find
DomainManager
Register Devices
and File System with
DomainManager
Figure 3.8: Device manager boot-up sequence.
different Internet Protocol (IP) address, but that general rule does not have to
apply to every possible implementation.
Application Control
Two classes in the framework control section of the SCA provide application
(waveform) control: Application and ApplicationFactory. ApplicationFactory is
the entity that creates waveforms; as such, ApplicationFactory contains a single
function, create().
ApplicationFactory
The create() function in the ApplicationFactory is called directly by the user, or
the cognition engine in the case of a CR system, to create a specific waveform.
The waveform’s individual components and their relative connections are
described in an eXtensible Markup Language (XML) file, and the component’s
specific implementation details are described in another XML file. The different
XML files that are used to describe a waveform in the SCA are described in (see
section Profiles).
Figure 3.9 shows an outline of the behavior of the ApplicationFactory. As seen
in Figure 3.9, the ApplicationFactory receives the request from an external source.
Upon receiving this request, it creates an instance of the Application class (discussed next), essentially a handle for the waveform. After the creation of the
Application object, the ApplicationFactory allocates the hardware capacities necessary in all the relevant hardware, checks to see if the needed Resources already
exist and creates them (through an entity such as the ResourceFactory) when they
do not already exist, connects all the Resources, and informs the DomainManager
(a master controller discussed in see section System Control) that the waveform
was successfully created. The other steps in this creation process, such as initializing the Resources, are not included in this description for the sake of brevity. Once
the ApplicationFactory has completed the creation process, a waveform comprising a variety of connected components now exists and can be used by the system.
103
Chapter 3
User
Request
Request Profile
from
DomainManager
Instantiate Needed
Objects
(Resources)
Create
Application
Object
Connect
Resources
Allocate
Needed Device
Capacities
Inform
DomainManager for
Successful
Installation
Figure 3.9: Simplified ApplicationFactory create() call for the creation of a waveform.
Application Class
The ApplicationFactory returns to the caller an instance of the Application class.
The Application class is the handle that is used by the environment to keep track
of the application. The Application class inherits from the Resource class, and its
definition does not add any more function calls, just application identifiers and
descriptors. The instance of the Application class that is returned by the
ApplicationFactory is the object that the user would call start, stop, and
releaseObject on to start, stop, and terminate the application, respectively.
System Control
In order for the radio system to behave as a single system, a unifying point is necessary. The specific nature of the unifying point can be as simple as a central registry, or as sophisticated as an intelligent controller. In the SCA, this unifying
point is neither. The DomainManager, which is the focal point for the radio, is
such an entity, and its task is as a central registry of applications, hardware, and
capabilities. Beyond this registry operation, the DomainManager also serves as a
mount point for all the different pieces of the hardware’s system file. The
DomainManager also maintains the central update channels, keeping track of
changes in status for hardware and software and also informing the different system components of changes in system status.
The point at which the DomainManager is created can be considered to be the
system boot-up. Figure 3.10 shows this sequence. As seen in Figure 3.10, the
DomainManager first reads its own configuration file. After determining the different operating parameters, the DomainManager creates the central file system,
called a FileManager in the context of the SCA, reestablishes the previous configuration before shutdown, and waits for incoming requests. These requests can be
104
The Software Defined Radio as a Platform for Cognitive Radio
DeviceManagers associating with the DomainManager, or new Applications
launched by the ApplicationFactory.
Instantiate
DomainManager
and Read
Configuration
File
Create
FileManager and
Naming Context
Re-establish
Previous
Configuraion
Wait for
Requests
Boot-Up
Figure 3.10: DomainManager simplified boot-up sequence.
Putting It Together
In general, the boot-up sequence of an SCA radio is partitioned into two different
sets of steps, Domain boot-up, and one or more Device boot-ups. Once the platform has been installed (i.e., file system(s) and device(s)), the system is ready to
accept requests for new waveforms. These requests can arrive either from the user
or from some other entity, even a cognitive engine. Figure 3.11 shows a simplified
boot-up sequence for the different parts of the SCA radio.
Boot-Up
Instantiate
DomainManager
and Read
Configuration
File
Create
FileManager and
Naming Context
Re-establish
Previous
Configuration
Wait for
Events
Repeat for
All HW
Platforms
Power-up
User
Request
Instantiate
DeviceManager
and Read
Configuration
File
Request Profile
from
DomainManager
Install Devices
and FileSystem
Create
Application
Object
Find
DomainManager
Allocate
Needed Device
Capacities
Register Devices
and FileSystem
with DomainManager
Instantiate Needed
Objects
(Resources)
Connect
Resources
Inform
DomainManager of
Successful
Installation
Figure 3.11: Simplified boot-up sequence and waveform creation for SCA radio.
Profiles
The SCA is composed of a variety of profiles to describe the different aspects of
the system. The collection of all files describes, “the identity, capabilities, properties, inter-dependencies, and location of the hardware devices and software components that make up the system” [11] and is referred to as the Domain Profile.
105
Chapter 3
The Domain Profile is in XML, a language similar to the HyperText Markup
Language (HTML), which is used to associate values and other related information to tags. XML is relatively easy for a human to follow and, at the same time,
easy for a machine to process.
The relationships among the different profiles is shown in Figure 3.12. As seen
in this figure, there are seven file types: Device Configuration Descriptor (DCD),
DomainManager Configuration Descriptor (DMD), Software Assembly Descriptor
(SAD), Software Package Descriptor (SPD), Device Package Descriptor (DPD),
Software Component Descriptor (SCD), and Properties Descriptor (PRF). In addition, there is Profile Descriptor, a file containing a reference to a DCD, DMD,
SAD, or SPD.
Domain Profile
0…n
0…n
1
<<DTDElement>>
Device Configuaration Descriptor
DomainManager
Configuration Descriptor
1
<<DTDElement>>
Software Assembly Descriptor
1
1
1…n
1…n
<<DTDElement>>
Profile Descriptor
1
<<DTDElement>>
Software Package Descriptor
<<DTDElement>>
Profile Descriptor
1
0…1
0…n
<<DTDElement>>
Device Package Descriptor
0…n
0…n
<<DTDElement>>
Properties Descriptor
0…n
<<DTDElement>>
Software Component Descriptor
Figure 3.12: Files describing different aspects of an SCA radio.
The profiles have a hierarchical structure that is split into three main tracks, all
of which follow a similar pattern. In this pattern, there is an initial file for the
track. This initial file contains information about that particular track as well as
the names of other files describing other aspects of the track. These other files will
describe their aspect of the system and, when appropriate, will contain a reference
to another file, and so on. Three files point to the beginning of a track: DCD,
106
The Software Defined Radio as a Platform for Cognitive Radio
DMD, and SAD. The DCD is the first file that is read by the DeviceManager on
boot-up, the DMD is the first file that is read by the DomainManager on boot-up,
and the SAD is the first file that is read by the ApplicationFactory when asked to
create an application. Each of these files will contain not only information about
that specific component, but one or more (in the case of the DMD, only one) references to an SPD. One SPD in this list of references contains information about
that specific component. Any other SPD reference is for an additional component
in that system: in the case of the DCD, a proxy for hardware; in the case of the
SAD, another component in the waveform.
The SPD contains information about that specific component, such as implementation choices, and a link to both a PRF, which contains a list of properties for
that component, and the SCD, which contains a list of the interfaces supported by
that component. In the case of the DCD, an additional track exists to describe the
actual hardware, the DPD. The DPD, much like the SPD, contains information
about that specific component, in this case, the hardware side. The DPD, also mirroring the SPD, contains a reference to a PRF, which contains properties information concerning capacities for that particular piece of hardware.
Application Programming Interface
To increase the compatibility between components, it is important to standardize
the interfaces; this means standardizing the function names, variable names, and
variable types, as well as the functionality associated with each component. The
SCA contains specifications for an API to achieve this compatibility goal. The
structure of the SCA’s API is based on building blocks, where the interfaces are
designed in a sufficiently generic fashion such that they can be reused or combined to make more sophisticated interfaces.
There are two types of APIs in the SCA: real-time, shown as A in Figure 3.13,
and non-real-time, shown as B in Figure 3.13. Real-time information is both data
and time-sensitive control, whereas non-real-time is control information, such as
setup and configuration, which does not have the sensitivity of real-time functionality. An interesting aspect of the SCA, API is that it describes interaction between
different layers as would be described in the Open Systems Interconnection (OSI)
protocol stack.
The API does not include interfacing information for intralayer communications. This means that in an implementation, the communications between two
baseband processing blocks, such as a filter and an energy detector, would be outside the scope of the specifications. This limitation provides a limit to the level of
granularity that the SCA supports, at least at the level of the API. To increase the
107
Chapter 3
Network
Waveform
Application
A
A
LLC
B
MAC
B
Physical
B
LLC
B
A
A
A
I/O
B
External
Network
Connection
A Data and Real-Time Control
B Non-Real-Time Control, Setup and Initialization,
from Applications, Other Levels, User Interface
Figure 3.13: SCA API for interlayer communications (I/O input/output; MAC medium
access control; LLC logical link control).
granularity to a level comparable to that shown in the example in Section 3.2, an
additional API is required.
3.5 Applications
3.5.1 Application Software
The application level of the system is where the cognitive engine is likely to operate. At this level, the cognitive engine is just another application in the system.
The biggest difference between a cognitive application and a traditional application such as a Web browser is that the cognitive engine has some deep connections with the underlying baseband software. Unlike baseband systems, which are
generally light and mainly comprise highly optimized software, application-level
design has more leeway in design overhead, allowing the use of a rich and varied
set of solutions that sometimes add significant performance overhead. Beyond the
typical OSs, such as products by Microsoft or products of the Linux family,
and the typical toolkits, such as wxWindows for graphing, or OS-specific solutions such as Microsoft’s DCOM, there exist three technologies that can have a
108
The Software Defined Radio as a Platform for Cognitive Radio
significant impact on the application environment for CR: Java, Binary Runtime
Environment for Wireless (BREW), and Python.
Java
Java is an object-oriented language developed at Sun Microsystems that allows
the development of platform-independent software while also providing some significant housekeeping capabilities to the developer.
Fundamentally, Java is a compiled language that is interpreted in real-time by
a virtual machine (VM). The developer creates the Java code and compiles it into
bytecodes by the Java compiler, and the bytecodes are then interpreted by the Java
VM (JVM) running on the host processor. The JVM has been developed in some
language and compiled to run (as a regular application) on the native processor,
and hence should be reasonably efficient when it runs on the target processor. The
benefit from such an approach is that it provides the developer the ability to “write
once, run anywhere.”
Java also provides some housekeeping benefits that can best be defined as
dreamlike for an embedded programmer, the most prominent of which is garbage
collection. Garbage collection is the JVM’s ability to detect when memory is no
longer needed by its application and to deallocate that memory. This way, Java
guarantees that the program will have no memory leaks. In C or C development, especially using distributed software such as CORBA, memory management is a challenge, and memory leaks are sometimes difficult to find and resolve.
Beyond memory management, Java also has an extensive set of security features
that provide limits beyond those created by the developer. For example, Java can
limit the program’s ability to access native calls in the system, reducing the likelihood that malicious code will cause system problems.
Java has several editions, each tailored to a particular type of environment:
●
●
●
Java 2 Standard Edition (J2SE) is designed to run in desktop environment,
servers, and embedded systems. The embedded design of J2SE relates to
reduced overhead inherent to Java, such as reduced memory or reduced computing power.
Java 2 Enterprise Edition (J2EE) is designed for large, multitier applications as
well as for the development of and interactions with Web services.
Java 2 Micro Edition (J2ME) is designed specifically for embedded and mobile
applications. J2ME includes features that are useful in a wireless environment,
such as built-in protocols and more robust security features.
109
Chapter 3
Beyond these versions of Java, other features are available that can be useful
for the radio developer. For example, under J2ME, there is a JavaPhone API. This
API provides application-level support for telephony control, datagram messages
(unreliable service over IP), power management, application installation, user profile, and address book and calendar. The JavaPhone API coupled with existing
software components can reduce the development time associated with an application suite.
Java has been largely successful, especially in the infrastructure market. In
such an environment, any overhead that may be incurred by the use of a real-time
interpreter is overwhelmed by the benefit of using Java. For the embedded environment, the success of Java has been limited. There are three key problems with
Java: memory, processing overhead, and predictability.
●
●
●
Memory: Most efforts to date in Java for the embedded world revolve around
the reduction of the overall memory footprint. One such example is the
Connected Limited Device Configuration (CLDC). The CLDC provides a set of
API and a set of VM features for such constrained environments.
Processing Overhead: Common sense states that processing overhead
is a problem whose relevance is likely to decrease as semiconductor manufacturing technology improves. One of the interesting aspects of processing overhead is that if the minimum supported system data rates for communications
systems grows faster than processing power per milliwatt, then processing
overhead will become a more pressing problem. The advent of truly flexible
software-defined systems will bear out this issue and expose whether it is in
fact a problem.
Predictability: Predictability is a real-time concern that is key if the application
software is expected to support such functionality as link control. Given that the
executed code is interpreted by the JVM, asynchronous events such as memory
management can mean that the execution time between arbitrary points of execution in the program can vary. A delay in execution can be acceptable because
the design may be able to tolerate such changes. However, large variations in
this execution time can cause such problems as dropped calls. A Real-time
Extension Specification for Java exists that is meant to address this problem.
Java is a promising language that has met success in several arenas. Efforts
within the Java community to bring Java into the wireless world are ongoing and
are likely to provide a constantly improving product.
110
The Software Defined Radio as a Platform for Cognitive Radio
BREW
Qualcomm created the BREW for its CDMA phones. However, BREW is not
constrained to just Qualcomm phones. Instead, BREW is an environment
designed to run on any OS that simplifies development of wireless applications.
BREW supports C/C and Java, so it is also largely language independent.
BREW provides a functionality set that allows the development of graphics
and other application-level features. BREW is designed for the embedded environment, so the overall footprint and system requirements are necessarily small.
Furthermore, it is designed for the management of binaries, so it can download
and maintain a series of programs. This set of features is tailored for the handset
environment, where a user may download games and other applications to execute
on the handset but lacks the resources for more traditional storage and management structures.
Python
Python is an “interpreted, interactive, OOP language” [12]. Python is not an
embedded language and it is not designed for the mobile environment, but it can
play a significant role in wireless applications, as evidenced by its use by
GNURadio. What makes Python powerful is that it combines the benefits of
object-oriented design with the ease of an interpreted language.
With an interpreted language, it is relatively easy to create simple programs
that will support some basic functionality. Python goes beyond most interpreted
languages by adding the ability to interact with other system libraries. For example, by using Python one can easily write a windowed application through the use
of wxWidgets. Just as Python can interact with wxWidgets, it can interact with
modules written in C/C, making it readily extendable.
Python also provides memory management structures, especially for strings,
that make application development simpler. Finally, because Python is interpreted,
it is OS-independent. The system requires the addition of an interpreter to function; of course, if external dependencies exist in the Python software for features
such as graphics, the appropriate library also has to exist in the new system.
3.6 Development
Waveform development for sophisticated systems is a complex, multidesigner
effort and is well beyond the scope of this book. However, what can be
approached is the design of a simple waveform. In this case, the waveform is
111
Chapter 3
designed using the SCA, bringing several of the concepts described thus far into a
more concrete example. Before constructing a waveform, it is important to build a
component.
3.6.1 Component Development
Several structures are possible for a component because it is defined only in terms
of its interfaces and functionality. At the most basic level, it is possible to create a
component class that inherits from both Resource and Port. In such a case, the
component would function as a single thread and would be able to respond to only
one event at a time. A diagram of this simple component is seen in Figure 3.14.
Figure 3.14: Simple resource and
port component. The component
is both a resource and a port.
Resource
and
Port
A more sophisticated approach is to separate the output ports into separate
threads, each interfacing with the primary Resource through some queue. This
approach allows the creation of fan-out structures while at the same time maintaining a relatively simple request response structure. A diagram of this fan-out
structure is seen in Figure 3.15.
Figure 3.15: Fan-out
structure for a component,
the resource uses output
ports but is the only entry
point into the component.
Port
Resource
Port
One of the problems with the component structure shown in Figure 3.15 is that
it does not allow for input interfaces that have the same interface name; this limitation increases the difficulty in using the API described in the specifications. To
resolve this problem, a fan-in structure needs to be added to the system, even though
this creates another level of complexity to the implementation. A way to implement
this fan-in structure is to mimic the fan-out structure, and to place each input Port
as a separate thread with a data queue separating each of these threads with the
functional Resource. An example of this implementation is shown in Figure 3.16.
112
The Software Defined Radio as a Platform for Cognitive Radio
The approaches shown in Figures 3.14, 3.15, and 3.16 each present a different
structure for the same concept. The specific implementation decision for a component is up to the developer, and can be tailored to the specific implementation.
Because the components are described in terms of interfaces and functionality, it
is possible to mix and match the different structures, allowing the developer even
more flexibility.
Port
Port
Resource
Port
Port
Figure 3.16: Fan-in, fan-out structure for a component. The resource uses both input and
output ports.
3.6.2 Waveform Development
As an example of the waveform development, assume that a waveform is to be
created that splits processing into three parts: baseband processing (assigned to a
DSP), link processing (assigned to a GPP), and a user interface (assigned to the
same GPP). A diagram of this waveform is shown in Figure 3.17.
Application
(Auto-Generated)
Assembly
Controller
Baseband
Processing Control
Link Processing
User
Interface
Figure 3.17: Simple SCA application. This example is made up of three functional
components and one control component.
113
Chapter 3
The application shown in Figure 3.17 has several pieces that should be readily
recognized; the application was generated by the framework as a result of the
Application Factory. Furthermore, the three proxies representing the relevant processing for the radio are also shown. Since the baseband processing is performed
on a DSP, the component shown is a proxy for data transfer to and from the DSP.
Link processing and the user interface, however, are actually implemented on the
GPP. The assembly controller shown is part of the SCA, but not as a separate
class. The assembly controller provides control information to all the different
deployed components. Because the framework is generic, application-specific
information, such as which components to tell to start and stop, must somehow be
included. In the case of the SCA, this information is included in the assembly controller, a custom component whose sole job is to interface the application object to
the rest of the waveform.
The waveform described in Figure 3.17 does not describe real-time
constraints; it assumes that the real-time requirements of the system are met
through some other means. This missing piece is an aspect of the SCA that
is incomplete.
3.7 Cognitive Waveform Development
In the context of SDR, a cognitive engine is just another component of a waveform, so the issue in cognitive engine deployment becomes one of component
complexity. In the simplest case, a developer can choose to create components
that are very flexible, an example of which is seen in Figure 3.18. The flexible
baseband processor seen in this figure can respond to requests from the cognitive
engine and alter its functionality. A similar level of functionality is also available
in the link-processing component. The system shown in Figure 3.18 is a slight
modification of the system shown in Figure 3.17. The principal difference in this
system is the addition of a reverse-direction set of ports and changed functionality
for each component.
The principal problem with the structure shown in Figure 3.18 is that it places
all the complexity of the system onto each separate component. Furthermore, the
tie-in between the cognitive engine and the rest of the waveform risks the engine
implementation to be limited to this specific system. An alternate structure is to
create a whole waveform for which the only functionality is a cognitive engine, as
seen in Figure 3.19.
The cognitive engine waveform shown in Figure 3.19 has no link to a waveform. To create this link, a link is created between the ApplicationFactory and the
114
The Software Defined Radio as a Platform for Cognitive Radio
Application
(Auto-Generated)
Assembly
Controller
Flexible Baseband
Processing Control
Flexible
Link Processing
Cognitive
Engine
Figure 3.18: Simple cognitive waveform. The waveform performs both communications and
cognitive functionality.
Figure 3.19: Cognitive engine waveform. The
waveform performs only cognitive functionality.
It assumes communications functionality is
performed by other waveforms.
Application
(Auto-Generated)
Assembly
Controller
Cognitive
Engine
cognitive engine waveform. The cognitive engine can then request that the
ApplicationFactory launch new waveforms. These new waveforms perform the
actual communications work, which is evaluated by the functioning cognitive
engine. If performance on the new waveform’s communications link falls within
some determined parameter, the cognitive engine can terminate the existing communications link waveform and request a new waveform (with different operating
parameters or a different structure altogether) from the ApplicationFactory. This
structure is seen in Figure 3.20.
An aspect of Figure 3.20 that is apparent is that the Port structure evident at
the waveform level, as seen in Figure 3.19, scales up to interwaveform communications. Another aspect of this approach is that the cognitive waveform does not
have to be collocated with the communications waveform. As long as timing
115
Chapter 3
Figure 3.20: Multi-waveform cognitive
support. The stand-alone cognitive
waveform requests new waveforms from
the SCA ApplicationFactory.
ApplicationFactory
Cognitive
Waveform
Communications
Waveform
constraints are met, the cognitive waveform can be placed anywhere within the
network that has access to the system. This aspect of the deployment of the waveforms allows the concept of a CR within the SCA to easily extend to a cognitive
network, where a single cognitive engine can control multiple flexible radios, as
seen in Figure 3.21.
Cognitive
Waveform
Flexible
Radio
Network
Flexible
Radio
Flexible
Radio
Figure 3.21: SCA-enabled cognitive network composed of multiple cognitive nodes.
With SDR and an astute selection of processing and RF hardware on the part
of a developer, it is possible to create highly sophisticated systems that can operate well at a variety of scales, from simple single-chip mobile devices all the way
up to multitiered cognitive networks.
3.8 Summary
Although a SDR is not a necessary building block of a CR, the use of SDR in CR
can provide significant capabilities to the final system. An SDR implementation is
a system decision, in which the selection of both the underlying hardware composition and the software architecture are critical design aspects.
116
The Software Defined Radio as a Platform for Cognitive Radio
The selection of hardware composition for an SDR implementation requires
an evaluation of a variety of aspects, from the hardware’s ability to support the
required signals to other performance aspects, such as power consumption and silicon area. Traditional approaches can be used to estimate the needs at the RF and
data acquisition levels. At the processing stage, it is possible to create an estimate
of a processing platform’s ability to be able to support a particular set of signal
processing functions. With such an analysis, it is possible to establish the appropriate mix of general-purpose processors, DSPs, FPGAs, and CCMs for a particular set of signal processing needs.
In order to mimic the nature of a hardware-based radio, with components such
as mixers and amplifiers, CBP is a natural way to consider software for SDR. In
CBP, components are defined in terms of their interfaces and functionality. This
definition provides the developer with significant freedom on the specific structure
of that particular component.
Even though a developer may choose to use CBP for the design of an SDR
system, a substantial infrastructure is still needed to support SDR implementations. This infrastructure must provide basic services, such as the creation and
destruction of waveforms, as well as general system integration and maintenance.
The goal of a software architecture is to provide this underlying infrastructure.
The SCA is one such architecture. The SCA provides the means to create and
destroy waveforms, manage hardware and distributed file systems, and manage
the configuration of specific components.
Finally, beyond programming methodologies and architectures are the actual
languages that one can use for development of waveforms and the specific patterns that are chosen for the developed software. The various languages have different strengths and weaknesses. C and Java are the dominant languages in
SDR today. Python, a scripting language, has become increasingly popular in
SDR applications, and is likely to be an integral part of future SDR development.
Much like the language selection, a design pattern for a particular component
can have a dramatic effect on the capabilities of the final product. Design patterns
that focus on flexibility can be more readily applied to cognitive designs, from the
most basic node development all the way up to full cognitive networks.
References
[1] J. Neel, J.H. Reed and M. Robert, “A Formal Methodology for Estimating the
Feasible Processor Solution Space for a Software Radio,” in Proceedings of the
117
Chapter 3
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
2005 Software Defined Radio Technical Conference and Product Exposition,
November 2005, Orange Co, CA.
T. Budd, An Introduction to Object-Oriented Programming, Third Edition,
Addison-Wesley, 2001, Boston, MA.
M. Weisfeld, The Object-Oriented Thought Process, Second Edition, Sams, 2003,
Indianapolis, IN.
E. Gamma, R. Helm, R. Johnson and J. Vlissides, Design Patterns: Elements of
Reusable Object-Oriented Software, Addison-Wesley, 1995, Boston, MA.
A. Shalloway and J.R. Trott, Design Patterns Explained: A New Perspective on
Object-Oriented Design, Second Edition, Addison-Wesley, 2005, Boston, MA.
J. Kerievsky, Refatcoring to Patterns, Addison-Wesley, 2005, Boston, MA.
http://www.gnu.org/software/gnuradio/
http://ossie.mprg.org/
http://www.crc.ca/en/html/crc/home/research/satcom/rars/sdr/sdr
M. Henning and S. Vinoski, Advanced CORBA Programming with C,
Addison-Wesley, 1999, Boston, MA.
JTRS Joint Program Office, JTRS-5000, “Software Communications Architecture
Specification, SCA V3.0,” August 2004, San Deigo, CA.
http://www.python.org/
118
CHAPTER 4
Cognitive Radio: The
Technologies Required
John Polson
Bell Helicopter
Textron Inc.
Fort Worth, TX, USA
4.1 Introduction
Technology is never adopted for technology’s sake. For example, only hobbyists used
personal computers (PCs) until a spreadsheet program, a “killer application,” was
developed. Then business needs and the benefits of small computers became apparent
and drove PC technology into ubiquitous use. This led to the development of more
applications, such as word processors, e-mail, and more recently the World Wide Web
(WWW). Similar development is under way for wireless communication devices.
Reliable cellular telephony technology is now in widespread use, and new
applications are driving the industry. Where these applications go next is of paramount importance for product developers. Cognitive radio (CR) is the name
adopted to refer to technologies believed to enable some of the next major wireless applications. Processing resources and other critical enabling technologies for
wireless killer applications are now available.
This chapter presents a CR roadmap, including a discussion of CR technologies and applications. Section 4.2 presents a taxonomy of radio maturity, and
Sections 4.3 and 4.4 present more detailed discussions. Sections 4.5 and 4.6 are
about enabling and required technologies for CRs. They present three classes of
cognitive applications, one of which may be the next killer application for wireless devices. Conjectures regarding the development of CR are included in
Section 4.7 with arguments for their validity. Highlights of this chapter are
119
Chapter 4
discussed in the summary in Section 4.8, which emphasizes that the technologies
required for CR are presently available.
4.2 Radio Flexibility and Capability
More than 40 different types of military radios (not counting variants) are currently in operation. These radios have diverse characteristics; therefore, a large
number of examples can be drawn from the pool of military radios. This section
presents the continuum of radio technology leading to the software-defined radio
(SDR). Section 4.3 continues the continuum through to CR.
The first radios deployed in large numbers were “single-purpose” solutions.
They were capable of one type of communication (analog voice). Analog voice
communication is not particularly efficient for communicating information, so
data radios became desirable, and a generation of data-only radios was developed.
At this point, our discussion of software and radio systems begins. The fixed-point
solutions have been replaced with higher data rate and voice capable radios with
varying degrees of software integration. This design change has enabled interoperability, upgradability, and portability.
It will be clear from the long description of radios that follows that there have
been many additional capabilities realized in radios over their long history. SDRs
and even more advanced systems have the most capabilities, and additional functions are likely to be added over time.
4.2.1 Continuum of Radio Flexibility and Capability
Basing CR on an SDR platform is not a requirement, but it is a practical approach
at this time because SDR flexibility allows developers to modify existing
systems with little or no new hardware development, as well as to add cognitive
capabilities. The distinction of being a CR is bestowed when the level of software
sophistication has risen sufficiently to warrant this more colorful description.
Certain behaviors, discussed in this chapter, are needed for a radio to be
considered a CR.
Historically, radios have been fixed-point designs. As upgrades were desired
to increase capability, reduce life cycle costs, and so forth, software was added to
the system designs for increased flexibility. In 2000, the Federal Communications
Commission (FCC) adopted the following definition for software radios: “A communications device whose attributes and capabilities are developed and/or implemented in software” [1]. The culmination of this additional flexibility is an SDR
system, as software capable radios transitioned into software programmable
120
Cognitive Radio: The Technologies Required
radios and finally became SDRs. The next step along this path will yield aware
radios, adaptive radios, and finally CRs (see Figures 4.1 and 4.2).
Software
Capable
Radio
Software
Programmable
Radio
SoftwareDefined
Aware
Radio
Adaptive
Radio
Cognitive
Radio
Increasing Technology/Software Maturity
Medium
(State of the shelf)
Technology
High
(State of the art)
Figure 4.1: SDR technology continuum. As software sophistication increases, the radio
system capabilities can evolve to accommodate a much broader range of awareness,
adaptivity, and even the ability to learn.
CR
JTRS
SUO-SAS
MIDS
Falcon
Jaguar
MBITR
MBMMR
PSC-5
ICNIA
JTT
NTDR
DMR
Leprechaun
EPLRS
LST-5 SINCGARS ASIP
PRC-117
Low
(Mature)
ARC-210
VRC-99
ARC-220
ARC-164
PLRS
SINCGARS
WSC-3
SoftwareCapable
SoftwareProgrammable
SoftwareDefined
Cognitive
SDR Intensity
Figure 4.2: Examples of software radios. The most sophisticated software control and
applications are reaching cognitive levels (see Tables 4.1–4.3 for descriptions).1
4.2.2 Examples of Software Capable Radios
Several examples of software capable radios are shown in Figure 4.2 and detailed
in Table 4.1. The common characteristics of these radios are fixed modulation
capabilities, relatively small number of frequencies, limited data and data rate
capabilities, and finally the ability to handle data under software control.
1
Note that the radios that are part of Figure 4.2 and Tables 4.1–4.3 are based on the best available public information and are intended only to notionally indicate a distribution of some of the well-known
radios. For additional, accurate information about capabilities, contact the respective manufacturers.
121
Table 4.1: Selected software capable radios.
SINCGARS (Single-Channel Ground and
Airborne Radio System)
PLRS (Position Location Reporting
System)
122
A family of very high frequency—frequency modulation (VHFFM) radios that provides a primary means of command and
control. SINCGARS has frequency-hopping capability, and
certain US Air Force versions operate in other bands using
amplitude modulation (AM) waveforms. The SINCGARS
family of radios has the capability to transmit and receive voice
and tactical data, and record traffic on any of 2320 channels
(25 kHz) between 30 and 88 MHz. A SINCGARS with an
Internet controller is software capable [2].
A command-and-control aide that provides real-time, accurate,
three-dimensional (3D) positioning, location, and reporting
information for tactical commanders. The jam-resistant ultra
high-frequency (UHF) radio transceiver network automatically
exchanges canned messages that are used to geolocate unit
positions to 15-m accuracy. Commanders use PLRS for
situational awareness. PLRS employs a computer-controlled,
crypto-secured master station and an alternate master station
to ensure system survivability and continuity of operations.
The network, under master station management, automatically
uses user units as relays to achieve over-the-horizon (OTH)
transmission and to overcome close-in terrain obstructions to
line-of-sight (LOS) communications. When a rugged portable
computer (PC) is used with the user unit, it becomes a
mini-command-and-control station with a local area map along
with superimposed with position and identification (ID)
information. The computer interface and network control make
PLRS a software capable radio system [3].
AN/WSC-3
AN/ARC-164 HaveQuick II
123
AN/ARC-220
A Demand Assigned Multiple Access (DAMA) satellite
communication (SATCOM) terminal. It meets tight size and
weight integration requirements. This single-waveform radio
has a computer (Ethernet) interface and is software capable [4].
A LOS UHF-AM radio used for air-to-air, air-to-ground, and
ground-to-air communications. ARC-164 radios are deployed
on all US Army rotary wing aircraft and provide anti-jam,
secure communications links for joint task force missions in the
tactical air operations band. This radio operates as a singlechannel (25 kHz) or a frequency-hopping radio in the
225–399.975 MHz band. The aircraft radio transmits at 10 W
output power and can receive secure voice or data. The
ARC-164 data-handling capability makes it software capable. It
has an embedded electronic counter-countermeasure (ECCM)
anti-jam capability. The 243.000 MHz guard channel can be
monitored. One model of the AN/ARC-164 is interoperable
with SINCGARS [5].
The standard high-frequency (HF) (2.0–29.999 MHz) radio for
US Army aviation. It includes secure voice and data
communications on any of 280,000 frequencies. The system
uses software-realized DSP. Using MIL-STD-2217, software
updates can be made over a MIL-STD-1553B bus. ARC-220
data processing capabilities optionally include e-mail
applications. The ARC-220 is software capable. MIL-STD-148141A automatic link establishment (ALE) protocols improve
connectivity on behalf of ARC-220 users. ALE executes in
microprocessors and is particularly interesting as a cognitive
application because of its use of a database and its sounding of
(Continued)
Table 4.1: Selected software capable radios (Continued)
AN/VRC-99
124
LST-5E
the RF environment for sensing. GPS units may be interfaced
with the radio, providing geolocation information [6].
A secure digital network radio used for high data rate
applications. This 1.2–2.0 GHz broadband direct-sequence
spread-spectrum radio has National Security Agency (NSA)
certified high-assurance capabilities and provides users with 31
RF channels. It supports TDMA and FDMA (frequency division
multiple access). The networking capabilities are software
programmable. AN/VRC-99 provides digital battlespace users
with the bandwidth to support multimedia digital terminal
equipment (DTE) of either an army command-and-control
vehicle (C2V), army battle-command vehicle (BCV), or a
marine corps advanced amphibious assault vehicle (A3V) with
a single wideband secure waveform that supports voice,
workstation, data, and imagery with growth for video
requirements [7].
A software capable UHF tactical SATCOM/LOS transceiver
with embedded communications security (COMSEC). The
LST-5E provides the user with a single unit for high-grade
half-duplex secure voice and data over both wideband (25 kHz)
AM/FM and narrowband (5 kHz) 1200 bits per second (bps)
binary phase shift keying (BPSK) and 2400 bps BPSK. Current
applications for the LST-5E include manpack, vehicular,
airborne, shipborne, remote, and fixed stations. The terminal is
compatible with other AM or FM radios that operate in the
225–399.995 MHz frequency band [8].
AN/PRC-6725 or Leprechaun
MBITR (MultiBand Intra/Inter
Team Radio)
CSEL (Combat Survivor/Evader Locator)
125
MIDS (Multifunction Information
Distribution System)
A handheld or wearable tactical radio. It can be programmed
from a frequency fill device, laptop or PC, system base station,
or it can be cloned from another radio [9].
Provides voice communications between infantry soldiers. The
Land Warrior Squad Radio is a SINCGARS-compatible,
eight-channel radio. MBITR is a software capable design
based on the commercially available PRC-6745 Leprechaun
radio [10].
Provides UHF communications and location for the purposes of
Joint Search and Rescue Center operations. CSEL uses GPS
receivers for geolocation. Two-way, OTH, secure beaconing
through communication satellites allows rescue forces to locate,
authenticate, and communicate with survivors/evaders from
anywhere in the world. The handheld receivers are one segment
of the CSEL command-and-control system. The satellite-relay
base station, Joint Search and Rescue Center software suite, and
radio set adapter units interface with the UHF SATCOM
network to provide CSEL capability. Upgrades through software
loads make CSEL radios software capable [11].
A direct-sequence spread-spectrum, frequency-hopping,
anti-jam radio that supports the Link 16 protocol for
communication between aircraft. Operates in the band around
1 GHz. Originally called Joint Tactical Information Distribution
System (JTIDS), this radio waveform has been redeveloped on a
new hardware platform, and is now being converted to a
JTRS-compliant radio. When this conversion is completed it
will be an SDR. It is currently in various stages of development
and production as an interoperable waveform for Europe and the
United States [12].
Chapter 4
4.2.3 Examples of Software Programmable Radios
Several examples of software programmable radios are shown in Figure 4.2 and
detailed in Table 4.2. The common characteristics of these radios are their ability
to add new functionality through software changes and their advanced networking
capability.
4.2.4 Examples of SDR
Only a few fully SDR systems are available, as shown in Figure 4.2 and detailed
in Table 4.3. The common characteristic of SDR systems is complete adjustability
through software of all radio operating parameters.
Other products related to SDR include GNURadio and the Vanu Anywave™
Base Station. GNURadio is a free software toolkit that is available on the Internet.
It allows anyone to build a narrowband SDR. Using a Linux-based computer, a
radio frequency (RF) front-end and an analog-to-digital converter (ADC), one can
build a software-defined receiver. By adding a digital-to-analog converter (DAC)
and possibly a power amplifier, one can build a software-defined transmitter [25].
The Vanu Anywave™ Base Station is a software-defined system that uses commercial off-the-shelf (COTS) hardware and proprietary software to build a wireless cellular infrastructure. The goal is simultaneous support for multiple standards,
reduced operating expenses, scalability, and future proofing (cost-effective
migration) [26].
4.3 Aware, Adaptive, and CRs
Radios that sense all or part of their environment are considered aware systems.
Awareness may drive only a simple protocol decision or may provide network
information to maintain a radio’s status as aware. A radio must additionally
autonomously modify its operating parameters to be considered adaptive. This
may be accomplished via a protocol or programmed response. When a radio is
aware, adaptive, and learns, it is a CR [27]. Only cutting-edge research and
demonstration examples of aware, adaptive, or CRs are available currently.
4.3.1 Aware Radios
A voice radio inherently has sensing capabilities in both audio (microphone) and
RF (receiver) frequencies. When these sensors are used to gather environmental
information, it becomes an aware radio. The local RF spectrum may be sensed in
pursuit of channel estimates, interference, or signals of interest. Audio inputs may
126
Table 4.2: Selected examples of software programmable radios.
AN/ARC-210
Racal 25
127
SINCGARS ASIP
(Advanced System
Improvement Program)
Provides fully digital secure communications in tactical and air traffic
control (ATC) environments (30–400 MHz). Additionally, the radio provides
8.33 kHz channel spacing capability to increase the number of available
ATC frequencies, and provides for growth to new very-high-frequency
(VHF) data link modes. Currently provided functions are realized through
software and provide integrated communications systems with adaptability
for future requirements with little or no hardware changes. In other words,
the ARC-210 is software programmable. The ARC-210 is integrated into
aircraft and operated via MIL-STD-1553B data bus interfaces. Remote
control is also available for manual operation. This radio is interoperable
with SINCGARS (Single-Channel Ground and Airborne Radio System) and
HaveQuick radios [13].
A compliant Project 25 public safety radio operating in the 136–174 MHz
band. Its 5 W peak transmit power, rugged and submersible housing, and
digital voice with Digital Encryption Standard (DES) make it an advanced
radio. The radio uses DSP and flash memory architectures, and supports 12
and 16 kbps digital voice and data modes. Racal posts software upgrades on
a protected Internet site. Software programmability enables field upgrades
and is possible due to its DSP and flash memory-based architecture [14].
Interoperates with SINCGARS radios and enhance operational capability in
the Tactical Internet (TI) environment. The ASIP models reduce size and
weight, and provide further enhancements to operational capability in the TI
environment. SINCGARS ASIP radios are software programmable and provide improved data capability, improved FEC for low-speed data modes, a
GPS interface, and an Internet controller that allows them to interface with
EPLRS and Battlefield Functional Area host computers. The introduction of
Internet Protocol (IP) routing enables numerous software capabilities for
radio systems [2].
(Continued)
EPLRS (Enhanced Position
Location Reporting System)
AN/PRC-117F
128
Jaguar PRC-116
JTT (Joint Tactical Terminal)
An AN/TSQ-158 that transmits digital information in support of tactical
operations over a computer-controlled, TDMA communications network.
EPLRS provides two major functions: data distribution and position location
and reporting. EPLRS uses a frequency-hopping, spread-spectrum waveform
in the UHF band. The network architecture is robust and self-healing. Radio
firmware may be programmed from external devices. The peak data rate of
525 kbps supports application layer radio-to-radio interaction [15].
A software programmable, multiband, multimode radio (MBMMR). The
PRC-117F operates in the 30–512 MHz band. Embedded COMSEC, SATCOM (satellite communication), SINCGARS, HaveQuick, and ECCM capabilities are standard. Various software applications, such as file transfer,
Transmission Control Protocol (TCP) with IP, and digital voice, are included
in this software programmable radio [7].
In service in more than 30 nations, including the UK (Army) and US
(Navy). It is a frequency-hopping, software programmable radio with considerable ECCM capability, resisting jamming by constantly shifting hopsets
to unjammed frequencies. Security is further heightened by use of a scrambler. The Jaguar-V can also be used for data transmission at a rate of
16 kbps, and it may tolerate up to 50 radio nets, each with dozens of radios,
at once; if each net is frequency hopping in a different sequence, it will still
transmit to all of them [16].
A high-performance, software programmable radio. Its modular functionality is backward- and forward compatible with the Integrated Broadcast
Service (IBS). Using a software download, JTT can accept changes in format and protocol as IBS networks migrate to a common format. Subsequent
intelligence terminals require total programmability of frequency, waveform,
number of channels, communication security, and transmission security,
making subsequent terminals SDRs [17].
Chapter 4
Table 4.2: Selected examples of software programmable radios (Continued )
NTDR (Near-Term Digital
Radio) system
AN/PRC-138 Falcon
MBITR (MultiBand Intra/
Inter Team Radio)
Cognitive Radio: The Technologies Required
129
MBMMR (MultiBand,
MultiMode Radio)
A mobile packet data radio network that links Tactical Operations Centers
(TOCs) in a brigade area (up to 400 radios). The NTDR system provides
self-organizing, self-healing network capability. Network management terminals provide radio network management. The radios interface with emerging Army Battle Command System (ABCS) automated systems and support
large-scale networks in mobile operations with efficient routing software
that supports multicast operations. Over-the-air programmability eliminates
the need to send maintenance personnel to make frequency changes [18].
A manpack or vehicular-mounted HF and VHF radio set. Capabilities
include frequency coverage of 1.6–60 MHz in SSB/CW/AME (singlesideband, continuous wave, AM equivalent) and FM in the VHF band and
100 preset channels. In the data mode, the AN/PRC-138 offers a variety of
compatible modem waveforms that allows it to be integrated into existing
system architectures. Specific features include embedded encryption, ALE,
embedded high-performance HF data modems, improved power consumption management, and variable power output [19].
An AN/PSC-5D(C) that enhances interoperability among special operation
forces units. The MBMMR supports LOS and SATCOM voice and data in
six basic modes: LOS, Maritime, HaveQuick I/II, SINCGARS, SATCOM,
and DAMA. Additional features include embedded TI Range Extension and
Mixed Excitation Linear Prediction (MELP) voice coding. Over-the-air
rekeying (OTAR), extended 30 to 420 MHz band, MIL-STD-188-181B high
data rate in LOS communications and SATCOM, and MIL-STD-188-184
embedded advanced data controller are supported [20].
Designated AN/PRC-148, the MBITR provides AM/FM voice and data
communications in the 30–512 MHz band. Development of the MBITR is an
outgrowth of Racal’s work on DSP and flash memory. MBITR is JTRS SCA
2.2 compliant and does not require a JTRS waiver. The information security
(INFOSEC) capabilities are software programmable [21].
Chapter 4
Table 4.3: Selected examples of SDRs.
DMR (Digital Modular Radio)
JTRS (Joint Tactical Radio
System)
SUO-SAS (Small Unit
Operations—Situational
Awareness System)
A full SDR capable of interoperability
with tactical systems such as HF, DAMA,
HaveQuick, and SINCGARS, as well as data
link coverage for Link 4A and Link 11.
These systems are programmable and include
software-defined cryptographic functions. The
US Navy is committed to migrating DMR to
SCA compliance to allow the use of JTRS
JPO provided waveforms. The DMR may be
reconfigured completely via on-site or remote
programming over a dedicated LAN or wide
area network (WAN). The four full-duplex
programmable RF channels with coverage
from 2.0 MHz to 2.0 GHz require no change
in hardware to change waveforms or security.
The system is controlled, either locally or
across the network, by a Windows®-based
HMI [23].
A set of current radio procurements for fully
SDRs. These radio systems are characterized
by SCA compliance that specifies an operating environment that promotes waveform
portability. The JTRS JPO is procuring more
than 30 waveforms that will ultimately be executable on JTRS radio sets. System packaging
ranges from embeddable single-channel form
factors to vehicle-mounted multi-channel systems. JTRS radios are the current state-of-theart technology and have the highest level of
software sophistication ever embedded into
a radio [24].
Developed by DARPA to establish the
operational benefits of an integrated suite of
advanced communication, navigation, and situation awareness technologies. It served as
a mobile communications system for small
squads of soldiers operating in restrictive terrain. SUO-SAS provides a navigation function
utilizing RF ranging techniques and other
sensors to provide very high accuracy [22].
130
Cognitive Radio: The Technologies Required
be used for authentication or context estimates or even natural language understanding or aural human–machine interface (HMI) interactions. Added sensors
enable an aware radio to gather other information, such as chemical surroundings,
geolocation, time of day, biometric data, or even network quality of service
(QoS) measures. The key characteristic that raises a radio to the level of aware
is the consolidation of environmental information not required to perform simple
communications. Utilization of this information is not required for the radio
to be considered aware. There is no communication performance motivation
for developing this class of aware radios, and it is expected that this set will
be sparse.
One motivation for an aware radio is providing information to the user. As an
example, an aware radio may provide a pull-down menu of restaurants within a
user-defined radius. The radio may gather this information, in the future, from
low-power advertisement transmissions sent by businesses to attract customers.
A military application may be a situational awareness body of information that
includes a pre-defined set of radios and their relative positions. As an example, the
radios exchange global positioning system (GPS) coordinates in the background,
and the aware radios gather the information for the user and provide it on request.
The radio is not utilizing the information but is aware of the situation.
One example of an aware radio is the code division multiple access (CDMA)
based cellular system proposed by Chen et al. [28]. This system is aware of QoS
metrics and makes reservations of bandwidth to improve overall QoS. Another
example of an aware radio is the orthogonal frequency division multiplexing
(OFDM) based energy aware radio link control discussed by Bougard et al. [29].
4.3.2 Adaptive Radios
Frequency, instantaneous bandwidth, modulation scheme, error correction coding,
channel mitigation strategies such as equalizers or RAKE filters, system timing
(e.g., a time division multiple access [TDMA] structure), data rate (baud timing),
transmit power, and even filtering characteristics are operating parameters that
may be adapted. A frequency-hopped spread-spectrum radio is not considered
adaptive because once programmed for a hop sequence, it is not changed. A
frequency-hopping radio that changes hop pattern to reduce collisions may be
considered adaptive. A radio that supports multiple channel bandwidths is not
adaptive, but a radio that changes instantaneous bandwidth and/or system timing
parameters in response to offered network load may be considered adaptive. If a
radio modifies intermediate frequency (IF) filter characteristics in response to
channel characteristics, it may be considered adaptive. In other words, if a radio
131
Chapter 4
makes changes to its operating parameters, such as power level, modulation, frequency, and so on, it may be considered an adaptive radio.
At this time, two wireless products exhibit some degree of adaptation: the digital European cordless telephone (DECT) and 802.11a.
DECT can sense the local noise floor and interference of all the channels from
which it may choose. Based on this sensing capability, it chooses to use the carrier
frequencies that minimize its total interference. This feature is built into hardware,
however, and not learned or software adaptive; thus, DECT is not normally considered an adaptive radio.
802.11a has the ability to sense the bit error rate (BER) of its link, and to
adapt the modulation to a data rate and a corresponding forward error correction
(FEC) that set the BER to an acceptably low error rate for data applications.
Although this is adaptive modulation, 802.11 implementations generally are dedicated purpose-fixed application-specific integrated circuit (ASIC) chips, not software defined, and thus 802.11 is not normally considered to be an adaptive radio.
4.3.3 Cognitive Radios
A CR has the following characteristics: sensors creating awareness of the environment, actuators to interact with the environment, a model of the environment that
includes state or memory of observed events, a learning capability that helps to
select specific actions or adaptations to reach a performance goal, and some
degree of autonomy in action.
Since this level of sophisticated behavior may be a little unpredictable in early
deployments, and the consequences of “misbehavior” are high, regulators will
want to constrain a CR. The most popular suggestion to date for this constraint is
a regulatory policy engine that has machine-readable and interpretable policies.
Machine-readable policy-controlled radios are attractive for several reasons.
One feature is the ability to “try out” a policy and assess it for impacts. The
deployment may be controlled to a few radios on an experimental basis, so it is
possible to assess the observation and measurement of the behaviors. If the result
is undesirable, the policies may be removed quickly. This encourages rapid decisions by regulatory organizations. The policy-driven approach is also attractive
because spatially variant or even temporally variant regulations may be deployed.
As an example, when a radio is used in one country, it is subject to that country’s
regulations, and when the user carries it to a new country, the policy may be
reloaded to comply in the new jurisdiction. Also, if a band is available for use during a certain period but not during another, a machine-readable policy can realize
that behavior.
132
Cognitive Radio: The Technologies Required
The language being used to describe a CR is based on the assumption of a
smart agent model, with the following capabilities:2
●
●
●
●
●
●
●
Sensors creating awareness in the environment.
Actuators enabling interaction with the environment.
Memory and a model of the environment.
Learning and modeling of specific beneficial adaptations.
Specific performance goals.
Autonomy.
Constraint by policy and use of inference engine to make policy-constrained
decisions.
The first examples of CRs were modeled in the Defense Advanced Research
Projects Agency (DARPA) NeXt Generation (XG) radio development program.
These radios sense the spectrum environment, identify an unoccupied portion,
rendezvous multiple radios in the unoccupied band, communicate in that band,
and vacate the band if a legacy signal re-enters that band. These behaviors are
modified as the radio system learns more about the environment; the radios are
constrained by regulatory policies that are machine interpretable. The first demonstrations of these systems took place late in 2004 and in 2005 [30].
4.4 Comparison of Radio Capabilities and Properties
Table 4.4 summarizes the properties for the classes of advanced radios described in
the previous sections. Classes of radios have “fuzzy boundaries,” and the comparison shown in the table is broad. There are certain examples of radios that fall outside the suggestions in the table. A CR may demonstrate most of the properties
shown, but is not required to be absolutely reparameterizable. Note that the industry
consensus is that a CR is not required to be an SDR, even though it may demonstrate most of the properties of an SDR. However, there is also consensus that the
most likely path for development of CRs is through enabling SDR technology.
4.5 Available Technologies for CRs
The increased availability of SDR platforms is spurring developments in CR. The
necessary characteristics of an SDR required to implement a practical CR are
2
A smart agent also has the ability to not use some or all of the listed capabilities.
133
Table 4.4: Properties of advanced radio classes.
Radio property
Software
capable
radio
134
Frequency hopping
X
Automatic link establishment
X
(i.e., channel selection)
Programmable crypto
X
Networking capabilities
Multiple waveform
interoperability
In-the-field upgradable
Full software control of all
signal processing, crypto,
and networking functionality
QoS measuring/channel state
information gathering
Modification of radio
parameters as function of
sensor inputs
Learning about environment
Experimenting with different settings
Software
programmable
radio
Softwaredefined
radio
Aware radio
Adaptive
radio
Cognitive
radio
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
*
X
*
X
*
X
X
X
X
X
X
X
* The industry standards organizations are in the process of determining the details of what properties should be expected of aware,
adaptive, and CRs.
Cognitive Radio: The Technologies Required
excess computing resources, controllability of the system operating parameters,
affordability, and usable software development environments including standardized application programming interfaces (APIs). This section discusses some
additional technologies that are driving CR. Even though this is not a comprehensive list of driving technologies, it includes the most important ones.
4.5.1 Geolocation
Geolocation is an important CR enabling technology due to the wide range of
applications that may result from a radio being aware of its current location and
possibly being aware of its planned path and destination.
The GPS is a satellite-based system that uses the time difference of arrival
(TDoA) to geolocate a receiver. An overview of this system is presented in
Chapter 8. The resolution of GPS is approximately 100 m. GPS receivers typically
include a one-pulse-per-second signal that is Kalman filtered as it arrives at each
radio from each satellite, resulting in a high-resolution estimate of propagation
delay from each satellite regardless of position. By compensating each pulse for
the predicted propagation delay, the GPS receivers estimate time to approximately
340 nanoseconds (ns, or 109 seconds) of jitter [31].
In the absence of GPS signals, triangulation approaches may be used to geolocate a radio from cooperative or even non-cooperative emitters. Chapter 8
discusses the classical approaches of TDoA, time of arrival (ToA), and, if the
hardware supports it, angle of arrival (AOA). Multiple observations from multiple
positions are required to create an accurate location estimate. The circular error
probability (CEP) characterizes the estimate accuracy.
4.5.2 Spectrum Awareness/Frequency Occupancy
A radio that is aware of spectrum occupancy may exploit this information for its
own purposes, such as utilization of open channels on a non-interference basis.
Methods for measuring spectrum occupancy are discussed in Chapter 5.
A simple sensor resembles a spectrum analyzer. The differences are in quality
and speed. The CR application must consider the quality of the sensor in setting
parameters such as maximum time to vacate a channel upon use by an incumbent
signal. It ingests a band of interest and processes it to detect the presence of signals above the noise floor. The threshold of energy at which occupancy is declared
is a critical parameter. The detected energy is a function of the instantaneous
power, instantaneous bandwidth, and duty cycle. An unpredictable duty cycle is
expected. Spectrum occupancy is spatially variant, time variant, and subject to
135
Chapter 4
observational blockage (e.g., deep fading may yield a poor observation). Therefore,
a distributed approach to spectrum sensing is recommended.
The primary problem associated with spectrum awareness is the hidden node
problem. A lurking receiver (the best example is a television (TV) set) may be
subjected to interference and may not be able to inform the CR that its receiver is
experiencing interference. Regulators, spectrum owners, and developers of CR are
working to find robust solutions to the hidden node problem. Again, a cooperative
approach may help to mitigate some of the hidden node problems, but a cooperative approach will not necessarily eliminate the hidden node problem.
In addition to knowing the frequency and transmit activity properties of a
radio transmitter, it may also be desirable for the radio to be able to recognize the
waveform properties and determine the type of modulation, thereby allowing a
radio to request entry into a local network. Many articles have been published on
this topic, as well as a textbook by Azzouz and Nandi [32]. Once the modulation
is recognized, then the CR can choose the proper waveform and protocol stack to
use to request entry into the local network.
4.5.3 Biometrics
A CR can learn the identity of its user(s), enabled by one or more biometric sensors.
This knowledge, coupled with authentication goals, can prevent unauthorized
users from using the CR. Most radios have sensors (e.g., microphones) that may
be used in a biometric application. Voice print correlation is an extension to an
SDR that is achievable today. Requirements for quality of voice capture and signal
processing capacity are, of course, levied on the radio system. The source radio
can authenticate the user and add the known identity to the data stream. At the
destination end, decoded voice can be analyzed for the purposes of authentication.
Other biometric sensors can be used for CR authentication and access control
applications. Traditional handsets may be modified to capture the necessary inputs
for redundant biometric authentication. For example, cell phones recently have
been equipped with digital cameras. This sensor, coupled with facial recognition
software, may be used to authenticate a user. An iris scan or retina scan is also
possible. Figure 4.3 shows some of the potential sensors and their relative
strengths and weakness in terms of reliability and acceptability [33].
4.5.4 Time
Included in many contracts is the phrase “time is of the essence,” testament to the
criticality of prompt performance in most aspects of human interaction. Even a
136
Cognitive Radio: The Technologies Required
Biometrics in Order of Effectiveness
Biometrics in Order of Social Acceptability
1. Palm scan
2. Hand geometry
3. Iris scan
4. Retina scan
5. Fingerprint
6. Voiceprint
7. Facial scan
8. Signature dynamics
9. Keyboard dynamics
1. Iris scana
2. Keyboard dynamics
3. Signature dynamics
4. Voiceprintb
5. Facial scana
6. Fingerprintc
7. Palm scanc
8. Hand geometryc
9. Retina scana
aRequires a camera scanner
bUtilizes a copy of the voice input (low impact)
cRequires a sensor in the push-to-talk (PTT) hardware
Figure 4.3: Biometric sensors for CR authentication applications. Several biometric
measures are low impact in terms of user resistance for authentication
applications.
desktop computer has some idea about what time it is, what day it is, and even
knows how to utilize this information in a useful manner (date and time stamping
information). A radio that is ignorant of time has a serious handicap in terms of
learning how to interact and behave. Therefore, it is important for the CR to know
about time, dates, schedules, and deadlines.
Time-of-day information enables time division multiplexing on a coarsegrained basis, or even a fine-grained basis if the quality of the time is sufficiently
accurate. Time-of-day information may gate policies in and out. Additionally,
very fine knowledge of time may be used in geolocation applications.
GPS devices report time of day and provide a one-pulse-per-second signal.
The one-pulse-per-second signal is transmitted from satellites, but does not arrive
at every GPS receiver at the same time due to differences in path lengths. A properly designed receiver will assess the propagation delay from each satellite and
compensate each of these delays so that the one-pulse-per-second output is synchronous at all receivers with only a 340 ns jitter. This level of accuracy is adequate for many applications, such as policy gating and change of cryptographic
keys. Increased accuracy and lowered jitter may be accomplished through more
sophisticated circuitry.
The local oscillator in a radio system may be used to keep track of time of
day. The stability of these clocks is measured at approximately 106. These
clocks tend to drift over time, and in the course of a single day may accumulate up
to 90 ns of error. Atomic clocks have much greater stability (1011), but have traditionally been large and power hungry. Chip-scale atomic clocks have been
137
Chapter 4
demonstrated and are expected to make precision timing practical. This will
enable geolocation applications with lower CEPs.
4.5.5 Spatial Awareness or Situational Awareness
A very significant role for a CR may be viewed as a personal assistant. One of
its key missions is facilitating communication over wireless links. The opposite
mission is impeding communications when appropriate. As an example, most
people do not want to be disturbed while in church, in an important meeting, or in
a classroom. A CR could learn to classify its situation into “user interruptible”
and “user non-interruptible.” The radio accepting aural inputs can classify a longrunning exchange in which only one person is speaking at a time as a meeting or
classroom and autonomously put itself into vibration-only mode. If the radio
senses its primary user is speaking continuously or even 50 percent of the time,
it may autonomously turn off the vibration mode.
4.5.6 Software Technology
Software technology is a key component for CR development. This section discusses key software technologies that are enabling CR. These topics include policy engines, artificial intelligence (AI) techniques, advanced signal processing,
networking protocols, and the Joint Tactical Radio System (JTRS) Software
Communications Architecture (SCA).
Policy Engines
Radios are a regulated technology. A major intent of radio regulatory rules is to
reduce or avoid interference among users. Currently, rules regarding transmission
and reception are enumerated in spectrum policy as produced by various spectrum
authorities (usually in high-level, natural language). Regulators insist that even a
CR adhere to spectrum policies. To further complicate matters, a CR may be
expected to operate within different geopolitical regions and under different regulatory authorities with different rules. Therefore, CRs must be able to dynamically
update policy and select appropriate policy as a function of situation.
Spectrum policies relevant to a given radio may vary in several ways:
1. Policies may vary in time (e.g., time of day, date, and even regulations changing from time to time).
2. Policies may vary in space (e.g., radio and user traveling from one policy regulatory domain to another).
138
Cognitive Radio: The Technologies Required
3. A spectrum owner/leaser may impose policies that are more stringent than
those imposed by a regulatory authority.
4. The spectrum access privileges of the radio may change in response to a
change in radio user.
As a result, the number of different policy sets that apply to various modes and
environments grows in a combinatorial fashion. It is impractical to hard-code
discrete policy sets into radios to cover every case of interest. The accreditation of
each discrete policy set is a major challenge. SDRs, for example, would require
the maintenance of downloadable copies of software implementations of each
policy set for every radio platform of interest. This is a configuration management
problem.
A scalable expression and enforcement of policy is required. The complexity
of policy conformance accreditation for CRs and the desire for dynamic policy
lead to the conclusion that CRs must be able to read and interpret policy. Therefore,
a well-defined, accepted (meaning endorsed by an international standards body)
language framework is needed to express policy. For example, if an established
policy rule is constructed in the presence of other rules, the union of all policies is
applicable. This enables hierarchal policies and policy by exception. As an example, suppose the emission level in band A is X dBm, except for a sub-band A for
which the emission-level constraint is Y dBm if a Z KHz guard band is allowed
around legacy signals. Even this simple structure is multi-dimensional. Layers of
exceptions are complex. The policy engine must be able to constrain behavior
according to the intent of the machine-readable policy. An inference capability is
needed to interpret multiple rules simultaneously.
In the case of spectrum subleasing, policies must be delegated from the lessor
to the lessee, and a machine-readable policy may be delegated. When a CR
crosses a regulatory boundary, the appropriate policy must be enabled. Policies
may also be used by the system in a control function.
The policy should use accepted standard tools and languages because the policy engine must be able to access automatic interpretation to achieve the goals of
CR applications. Policies may be written by regulatory agencies or by third parties and approved by regulators, but in all cases policy is a legal or contractual
operating requirement and provability in the policy interpretation engine is needed
for certification.
Several suggestions for policy language have emerged. The eXtensible
Markup Language (XML) is not appropriate because it does not typically have
inference capabilities in the interpretation engines. The Ontology Inference Layer
139
Chapter 4
(OIL), Web Ontology Language (OWL), and DARPA Agent Markup Language
(DAML) have all been explored as possible policy languages. DARPA’s XG program
cites the OWL language as an appropriate language. Tool sets are available for
building policy definitions and for machine interpretation of the definitions [34].
AI Techniques
The field of AI has received a great deal of attention for decades. In 1950, Alan
Turing proposed the Turing test, regarding interacting with an entity and not being
able to tell if it is human or machine. The AI techniques that work are plentiful,
but most are not widely applicable to a wide range of problems. The powerful
techniques may even require customization to work on a particular problem.
An agent is an entity that perceives and acts. A smart agent model is appropriate for CR. Figure 4.4 explains four models of smart agents—simple reflex agents,
Simple Reflex Agent
Reflex Agent with State
AKA Model-Based Reflex Agents
• Agent selects its action as a function of the
current percept
• Combinational logic
•
•
Agent maintains internal state (memory) as a function of
percept history and partially reflects unobserved aspects
If it is not just a set of percepts, it is “modeling” the
environment
Sensors
Actuators
Map:
(Sensors, Memory,
Simple Model) -> Actions
Agent
Goal-Based Agent
• Utility function maps state sequence to real number
• Real number is the relative happiness of the agent
• Allows rational decisions in some cases where goals
are inadequate
Sensors
Agent
Environment
Actuators
Sensors
Map:
(Sensors, Memory,
Realistic Model, with
Feedback) ->
Sequence of Actions
Environment
Agent
Actuators
Utility-Based Agent
• Goal information identifies states that are desirable
• Simple: single step reaches goal(s)
• Complex: sequence of steps required to reach
goal(s)
Map:
(Sensors, Memory,
Realistic Model) ->
Sequence of Actions
Environment
Agent
Environment
Simple Map:
(Sensors) -> Actions
Sensors
Actuators
Figure 4.4: Four smart agent models. Smart agents provide a framework that is consistent
with the continuum of software maturity in software radios.
140
Cognitive Radio: The Technologies Required
model-based reflex agents, goal-based agents, and utility-based agents—defined
as follows:
1. A simple reflex agent is a simple mapping from current sensor inputs to actuator settings. This is a stateless agent model that neither learns nor adapts to the
environment.
2. A model-based reflex agent is still a simple mapping but now includes memory
of past inputs. The actions are a function of the current sensor inputs and the
recent past inputs, making it a finite memory agent model. There is still no
learning. Adaptation is limited, but this is the minimum level of sophistication
for an adaptable radio.
3. A goal-based agent adds to the memory of past inputs; it is a “realistic” model
of the environment. Now a sequence of actions may be “tested” against a goal
and an appropriate next action may be selected. The level of sophistication for
the model of the environment is not well defined. These agents have increased
capability of adapting because a prediction about the consequences of an
action is available. There is no feedback, and learning is therefore limited. This
is the minimal level of sophistication for a CR.
4. A utility-based agent maps the state sequence (memory state) to a “happiness”
value and therefore includes feedback. The more sophisticated environment
model may experiment with sequences of actions for selection of the best next
action. This model of a CR has the ability to learn and adapt [35].
A smart agent model for CR is appropriate. The agent framework supports the
continuum of radio maturity, and it allows the modular introduction of various AI
techniques from fuzzy control to genetic algorithms (GAs). Agents may be tailored to the application’s environment. In this sense, environment may be characterized in the following dimensions: fully observable versus partially observable,
deterministic versus stochastic, episodic versus sequential, static versus dynamic,
discrete versus continuous, and single agent versus multiagent.
The following is an incomplete list of AI techniques likely to find applicability
to CR, and further described in subsequent chapters:
●
●
●
●
State space models and searching
Ontological engineering
Neural networks in CR
Fuzzy control in CR
141
Chapter 4
●
●
●
GAs in CR
Game theory in CR
Knowledge-based reasoning in CR
Signal Processing
Digital signal processing (DSP) technology enables rapid advances in CRs.
Intellectual property resources are widely available for signal processing functions. In GPP (general-purpose processor) or DSP resources, libraries of routines
realize functions in efficient assembly language. In FPGA (field programmable
gate array) or ASIC resources, licensable cores of signal processing engines are
available in very-high-speed integrated circuit (VHSIC) Hardware Design
Language (VHDL) or Verilog. Signal processing routines are available for communication signal processing (modulation/demodulation, FEC, equalization,
filtering, and others); audio signal processing (voice coding, voice generation,
natural language processing); and sensor signal processing (video, seismic,
chemical, biometric, and others).
Synthesizing signal processing functions together to form a system is a complex task. A process for algorithm development, test case generation, realization,
verification, and validation eases the process of building a waveform or a cognitive system. Integrated tools for system development are available. Many of the
tool sets include automatic generation of high-level language source code or hardware definition language code. A bit-level accurate simulation environment is used
to develop the system algorithms and to generate test cases for post-integration
verification and validation. This environment may be used for CRs to synthesize
communications or multimission waveforms that enable a CR to achieve
specific goals.
Networking Protocols
Cooperative groups (a multiagent model) have the potential to increase capabilities in a variety of ways. For example, a lone CR is limited in its ability to access
spectrum, but a pair of CRs can sense the spectrum, identify an unused band, rendezvous there, and communicate. A network of CRs enables other significant
increases in capabilities. Software for Mobile Ad hoc Networking (MANET),
although maturing slowly, is a key enabling technology.
The medium access control (MAC) layer is critical in CR networks. If the CR
is employing advanced spectrum access techniques, a robust MAC that mitigates
142
Cognitive Radio: The Technologies Required
the hidden node problem is needed. In a “static spectrum access” environment, a
more traditional MAC is possible. A carrier sense collision detection (802.11
MAC) mode is not possible because a radio cannot receive in the exact same band
in which it is transmitting, so a carrier sense collision avoidance approach is frequently used. Request-to-send (RTS) and clear-to-send (CTS) messaging are popular wireless MACs amenable to MANETs. Other approaches include TDMA or
CDMA MACs.
The architecture for routing packets is important for performance in MANETs.
The approaches are generally divided into proactive and reactive algorithms. In a
proactive-routing environment, routing data, often in the form of a routing table,
are maintained so that a node has a good idea of where to send a packet to
advance it toward its final destination, and a node may know with great confidence how to route a packet even if one is not ready to go. Maintaining this
knowledge across a MANET requires resources. If the connection links are very
dynamic or the mobility of the nodes causes rapid handoff from one “network” to
another, then the overhead to maintain the routing state may be high. In contrast,
reactive-routing approaches broadcast a short search packet that locates one or
more routes to the destination and returns that path to the source node. Then the
information packet is sent to the destination on that discovered route. This causes
overhead in the form of search packets. Proactive and reactive routing both have
pros and cons associated with their performance measures, such as reliability,
latency, overhead required, and so on. A hybrid approach is often best to provide
scalability with offered network load.
An interesting application of CR is the ability to learn how to network with
other CRs and adapt behavior to achieve some QoS goal such as data rate
below some BER bound, bounded latency, limited jitter, and so forth. Various
cognitive-level control algorithms may be employed to achieve these results.
As an example, a fixed-length control word may be used to parameterize a communications waveform with frequency, FEC, modulation, and other measures.
The deployment of a parameterized waveform may be controlled and adapted
by using a generic algorithm and various QoS measures to retain or discard a
generated waveform.
Software Communications Architecture
The primary motivations for SDR technology are lower life cycle costs and
increased interoperability. The basic hardware for SDR is more expensive than for
a single-point radio system, but a single piece of hardware performs as many
radios. The single piece of hardware requires only one logistics tail for service,
143
Chapter 4
training, replacement parts, and so on. One of the driving costs in SDR development is that of software development. The JTRS acquisitions are controlling these
costs by ensuring software reuse. The approach for reuse is based on the Software
Communications Architecture (SCA), which is a set of standards that describes the
software environment. It is currently in release 2.2.1. Software written to be SCA
compliant is more easily ported from one JTRS radio to another. The waveforms
are maintained in a JTRS Joint Program Office (JPO) library.
A CR can be implemented under the SCA standards. Applications that raise
the radio to the level of a CR can be integrated in a standard way. It is expected
that DARPA’s XG program will provide a CR application for policy-driven,
dynamic spectrum access on a non-interference basis that executes on JTRS
radios. XG is the front-runner in the race to provide the first military CR.
4.5.7 Spectrum Awareness and Potential for Sublease or Borrow
The Spectrum Policy Task Force (SPTF) recommends that license holders in
exclusive management policy bands be allowed to sublease their spectrum. Figure
4.5 shows a sequence diagram for spectrum subleasing from a public safety spectrum owner. During the initial contact between the service provider and the public
safety spectrum manager, authentication is required. This ensures that spectrum
use will be accomplished according to acceptable behaviors and that the bill will
be paid [36].
For a subleasing capability to exist in a public safety band a shut-down-oncommand function must be supported with a bounded response time. There are
three approaches to this: continuous spectrum granting beacon, time-based granting of spectrum, and RTS–CTS–inhibit send. Figure 4.5 shows the time-based
granting of spectrum, described as a periodic handshake confirming sublease.
Even though public safety organizations may not sublease spectrum, other
organizations may choose to do so. Subleasing has the benefit of producing an
income stream from only managing the resource. Given proper behavior by
lessees and lessors, the system may become popular and open up the spectrum for
greater utilization.
4.6 Funding and Research in CRs
DARPA is funding a number of cognitive science applications, including: the XG
Program, Adaptive Cognition-Enhanced Radio Teams (ACERT), Disruption
Tolerant Networking (DTN), Architectures for Cognitive Information Processing
(ACIP), Real World Reasoning (REAL), and DAML. DARPA research dollars
144
Cognitive Radio: The Technologies Required
Subscriber 1
Subscriber 2
Service
Provider
Public Safety
Spectrum Manager
Subscriber 1 Calls Subscriber 2
Subscriber 1 Talks with Subscriber 2 on Usual Channel
Service Provider Running Out of Channels
Service Provider Requests Sublease
Spectrum Manager Grants Sublease
Service Provider Commands Channel Change
Subscriber 1 Talks with Subscriber 2 on Sublease
Periodic Handshake Confirming Sublease
Service Provider Commands Channel Change on Sublease Termination
Termination Acknowledge
Billing Record Sent
Figure 4.5: Spectrum subleasing sequence. One motivation for CR is the potential income
stream derived from subleasing idle spectrum on a non-interference basis.
under contract are easily in the tens of millions. Good results have been achieved
in many of these efforts.
The National Science Foundation (NSF) is also funding cognitive science
applications including grants to the Virginia Polytechnic Institute and State
University (Virginia Tech or VT). Additionally, NSF has sponsored Information
Theory and Computer Science Interface workshops that communicate CR
research results.
The SDR Forum has a CR Working Group that is investigating various CR
technologies such as spectrum access techniques, and a Cognitive Applications
Special Interest Group that is working with regulatory bodies, spectrum owners,
and users to communicate the potential of CR technologies. Numerous
145
Chapter 4
organizations participate at SDR Forum meetings. It is expected that the SDR
Forum will solicit inputs from industry through a Request for Information (RFI)
process in the near future.
Both the Federal Communications Commission (FCC) and the National
Telecommunications and Information Administration (NTIA) have interest in CR.
The FCC has solicited various comments on rule changes and an SPTF report.
NTIA has been involved in the discussions as they relate to government use of
spectrum.
4.6.1 Cognitive Geolocation Applications
If a CR knows where it is in the world, myriad applications become possible. The
following is a short set of examples. Figure 4.6 shows a use case level context diagram for a CR with geolocation knowledge.
Figure 4.6: CR use case for
geolocation. Several interactions
between the CR and the world are
enabled or required by geolocation
information.
User
Spatially Variant
Advertisers
GPS Satellites
CRs Geolocation
Spatially Aware
Routing
Boundary Aware
Policy Deployment
Space and Time
Aware Scheduling
of Tasks
Establishing the location of a CR enables many new functions. A cognitive
engine for learning and adapting can utilize some of the new functions. There are
multiple methods for determining location. For example, GPS receiver technology
is small and inexpensive. Given an appropriate API, numerous applications
including network localization (discussed next) and boundary awareness are available to the cognitive engine in the CR. Other methods for determining location are
discussed in Chapter 8.
Network localization is a term to describe position aware networking enhancement. For example, if a radio knows it is in a car, that it is morning rush hour, and
that the path being taken is the same as that of the last 200 days of commuting, it
146
Cognitive Radio: The Technologies Required
can predict being in the vicinity of the owner’s office wireless local area network
(LAN) within a certain time period. The radio may wait until it is in the office to
download e-mail and upload the pictures of the accident taken a few minutes ago.
This is an example of the highest level of management of radio functions (assuming the radio is used for e-mail, photos, etc.).
Spatial awareness may be used for energy savings. In a multiple short hoprouting algorithm with power management just closing the wireless link with
multiple short hops is usually more energy efficient than one long hop. Knowing
the position of each node in a potential route allows the CR to take energy consumption into consideration when routing a packet.
Spatially variant searching is a powerful concept for increasing a user’s operating efficiency. If it is time for supper, the CR may begin a search for the types of
restaurants the user frequents and sort them by distance and popularity. Other spatially variant searches are possible.
A radio aware of boundaries may be able to invoke policy as a function of
geopolitical region. When passing from one regulatory jurisdiction to another, the
rules change. The radio can adopt a conservative operation mode near the boundaries and change when they are crossed. The radio must have the ability to distinguish one region from another. This may require a database (which must be kept
up-to-date) or some network connectivity with a boundary server. Figure 4.7
shows a simplified sequence diagram in which a CR accesses spectrum as a function of a spatially variant regulatory policy.
A sequence of position estimates may be used to estimate velocity. Take a scenario in which teenagers’ cell phones would tattle to their parents when their
speed exceeds a certain speed. For example, suppose tattle mode is set. The radio
is moving at 45 mph at 7:30 a.m. The radio calls the parent and asks a couple of
questions such as: “Is 45 mph okay at this time? Is time relevant?” The questions
will be in the same order each time and the parent won’t have to wait for the
whole question, just the velocity being reported. Then the parent keys in a
response. After a few reports, the radio will develop a threshold as a function of
time. For example, during the time the radio is heading for school, 45 mph is okay.
During the lunch break (assuming a closed campus), 15 mph might be the threshold. An initial profile may be programmed, or the profile may be learned through
tattling and feedback to the reports. Vehicle position and velocity might also be
useful after curfew.
The CR application uses special hardware or customized waveforms that
return geolocation information. This information is used to access databases of
policies or resources to make better decisions. Dynamic exchange of information
147
Chapter 4
Location
Function
Cognitive
Engine
Policy
Engine
Boundary
and Resources
Function
Where Am I?
Position
Inquire Region (Position)
Region
Decision Regarding
Band to Use
Inquire (Band, Region,Time)
Approval
CR Accesses Band
Figure 4.7: Spectrum access sequence diagram. A policy engine uses regional inputs to select
spatially variant policies to approve or disapprove a requested spectrum access.
may be used for other networking actions. The set of CR applications that are
enabled by geolocation capability is large and has many attractive benefits.
4.6.2 Dynamic Spectrum Access and Spectrum Awareness
One of the most common capabilities of CRs is the ability to intelligently utilize
available spectrum based on awareness of actual activity. Current conservative
spectrum management methods (static spectrum assignments) are limited because
they reduce spatial reuse, preclude opportunistic utilization, and delay wireless
communication network deployment. Without the need to statically allocate spectrum for each use, however, networks can be deployed more rapidly. A CR with
spectrum sensing capability and cooperative opportunistic frequency selection is
an enabling technology for faster deployment and increased spatial reuse.
Spectrum access is primarily limited by regulatory constraints. Recent measurements show that spectrum occupancy is low when examined as a function of
frequency, time, and space [36]. CRs may sense the local spectrum utilization
148
Cognitive Radio: The Technologies Required
either through a dedicated sensor or by using a configured SDR receiver channel.
Uses of this information may create increased spectrum access opportunities.
One of the primary considerations for such a cognitive application is noninterference with other spectral uses. Figure 4.8 shows local spectrum awareness
and utilization. If the regulatory body is allowing CRs to utilize the unoccupied
“white space,” increased spectral access can be achieved. The CR can examine the
signals and may extract detailed information regarding use. By estimating the
other uses and monitoring for interference, two CRs may rendezvous at an unoccupied “channel” and communicate.
900,000
Frequency (MHz)
Spectrally Aware
Channel Activity Statistics
Usage Policies
– Lockouts, Rentals, Unlicensed
880,000
860,000
840,000
6781.000
Regulators Will Define
Spectrum Blocks Subject to
CR Commons
Etiquette and Who Is Allowed
to Use
6784.000
6787.000
Time of Day (Seconds)
6790.000
7–14% Occupied
Spectrum
Deployed Waveforms
Existing Signals
Detailed Signal Parameters
Figure 4.8: Spectrum awareness. A CR, or a set of CRs, may be aware of the spectrum and
may exploit unoccupied spectrum for its own purposes.
Sophisticated waveforms that have the ability to periodically stop transmitting
and listen for legacy users are called for in this application, as well as waveforms
that can adapt their spectral shape over time. Dynamic selection of channels to
utilize or vacate is important. Simulations of these cooperating CRs already exist,
and additional field demonstrations are expected in the near future. Another
advantageous waveform characteristic is discontinuous spectrum occupancy. This
allows a wideband communication system to aggregate available spectrum
149
Chapter 4
between other existing signals. Careful analysis is needed to ensure that sufficient
guard bands are utilized.
Figure 4.9 shows five suggested alternatives for utilizing spectrum in and
around legacy signals. The characteristics of the legacy signals may be provided
to the cooperating CRs by federated sensors. An alternative method for characterizing the legacy signals is time division sharing of a channel as a sensor and providing a look-through capability by duty-cycling transmit and monitor functions.
The five methods shown in Figure 4.9 avoid the legacy signal in various ways.
Frequency
Occupied
Frequency Spaces
b
1. Move in
Frequency
Frequency
d
c
a
2. Adapt BW and
Frequency
Time
Frequency
Time
5. Dynamic
Frequency
Hop Spread
Spectrum
Frequency
b
3. Dynamic
OFDM
a
c
d
4. Dynamic Direct
Spreading Spread
Spectrum
Time
Frequency
Frequency
c
d
e
d
e
b
a
b
c
d
e
a
c
b
Time
Time
c
Time
Figure 4.9: Non-interference methods for dynamic spectrum access. Different strategies for
deploying non-interfering waveforms have been proposed.
The first method assumes a fixed bandwidth waveform: the center frequency
may be adapted. The second method assumes a variable bandwidth signal, such as
a direct-sequence spread-spectrum waveform, where the chip rate is adapted and
the center frequency is adapted. The third method uses a water-filling method to
populate a subset of the carriers in an OFDM waveform. These three methods
150
Cognitive Radio: The Technologies Required
impact legacy signals very little if appropriate guard bands are observed. The
fourth method is a direct-sequence spread-spectrum waveform that underlies the
legacy signals. The interference from this underlay must be very small so that
legacy systems do not experience noise; thus the processing gain of the spread
spectrum underlay must be very high. The last method shown also avoids the
legacy signals by frequency hopping into only unoccupied channels.
The spread-spectrum method deserves some elaboration at this point. Legacy
receivers will perceive the spread-spectrum signal as an increase in the noise floor,
which may result in a decrease in the link margin of the legacy signal. The spreadspectrum receiver will perceive the legacy waveform(s) as a narrowband interference. The de-spreading of the desired signal will cause the narrowband signals to
spread. This spreading will cause the narrowband signals to appear at a signal
level reduced in power by the spreading gain, and thus appear as noise to the CR,
resulting in reduced link margin. The ability of each of these communication systems to tolerate this reduced link margin is link specific and therefore a subject of
great concern to the legacy system operators.
An OFDM waveform, the third method in Figure 4.9, has several benefits
including flat fading subchannels, elimination of equalizer due to long symbol
time and guard intervals, the ability to occupy a variety of bandwidths that “fit to
the available opportunity,” and the ability to null subcarriers to mitigate interference. Variable bit loading enables pre-nulling and dynamic nulling. Table 4.5
compares several methods for dynamic bit loading an OFDM waveform. Not
loading subcarriers occupied by legacy signals with guard bands around them
minimizes interference between CR and non-CR systems [37–39].
Because spectrum utilization is a spatially and temporally variant phenomenon, it requires repeated monitoring and needs cooperative, distributed coordination. The familiar hidden node and exposed node problems have to be considered.
Figure 4.10 shows a context diagram in which external sensor reports are made
available to the CR and may be considered when selecting unoccupied bands.
Figure 4.11 shows a sequence diagram in which a set of CRs is exchanging
sensor reports and is learning about local spectrum occupancy. At some time, a
pair of CRs wishes to communicate and rendezvous at a band for that purpose. When
a legacy signal is detected, the pair of CRs must vacate that band and relocate to
another.
The sensor technology utilized for spectrum awareness should be of high quality to mitigate the hidden node problem. For example, if a CR wants to use a TV
band, its sensor should be significantly more sensitive than a TV set so that if it
detects a TV signal, it will not interfere with local TV set reception, and if it does
151
Chapter 4
Table 4.5: Comparison of variable bit loading algorithms.
Method
Characteristic
Complexity
Water filling
Original approach, optimal,
frequently used for comparison
Optimal, loads bits serially
based on subcarrier energy
level, slow to converge,
repeated sorts
Suboptimal, rounds to integer
rates using signal-to-noise
gap approximations, some
sorting required
Optimal, computationally
efficient, efficient table lookup,
Lagrange multiplier with
bisection search, integer bit
loading, power allocation,
fast convergence
O(N2)
Hughes-Hartogs [37]
Chow [38]
Lagrange
(unconstrained)
Krongold [39]
O(SN2)
O(N log N N log S)
O(N log N),
revised O(N)
Note: N: number of subcarriers; S: number of bits per subcarrier.
User
Spectrum Usage
Policy
Other CR Sensor
Reports
CR Spectrum
Access
“Opportunistic”
Use of Spectrum
(with Look-through)
Sensor
Reports
Rendezvous
Protocol
Figure 4.10: Spectrum access context diagram. A spatially diverse sensing protocol is
required to mitigate such problems as hidden node or deeply fading RF channels for CR
accessto spectrum on a non-interference basis.
not detect a signal, there is a high probability that no TV set is near enough to
demodulate a TV signal on that channel.
Dynamic spectrum access as a result of learning about the spectrum occupancy is a strong candidate for a CR application. Numerous discussions in regulatory organizations have involved whether to allow this behavior. If implemented
152
Cognitive Radio: The Technologies Required
Remote CR
(Talk with)
Local CR
(Talks with)
Other CR
(No Communications)
Legacy
Users
Spectrum Usage
Sensor Report
Sensor Report
Sensor Report
Sensor Report
Request Communication
Decision Regarding
Band to Use
Decision Regarding
Band to Use
Rendezvous Protocol
Communications
Spectrum Usage
Sensor Report
Move Band
Communications
Figure 4.11: Dynamic spectrum access sequence diagram (assumes a control channel).
A pair of CR communicating on an unoccupied channel must vacate when interference
potential is detected.
correctly, it is a win–win situation in which more spectrum is utilized and very little additional interference is suffered.
4.6.3 The Rendezvous Problem
The difficulty for the CR network is the radios locating each other and getting the
network started. Before any transmission can occur, all radios must survey the
spectrum to determine where the available “holes” (frequency reuse opportunities)
are located. However, each receiver–sensor will perceive the spectrum slightly differently because each will see different objects shadowing different transmitters,
and will either see each transmitter with a different signal strength or will not see
some transmitters that other nodes are able to see. So while node A may see an
open frequency, node B may consider that frequency to be in use by a distant
transmitter node or network L. To get the network started, the CRs must agree on
a protocol to find each other.
153
Chapter 4
There are several possible methods by which to do this. These methods
depend strongly on whether there is an infrastructure in place to help start up a
CR network, or whether no infrastructure can be assumed.
Infrastructure-Aided Rendezvous
If we assume that there is an infrastructure component, we must also assume it
behaves just like the CR and does not interfere with legacy systems. We assume,
however, that it periodically transmits a beacon signal, and we assume that this beacon includes reference time, next frequency-hop(s), and a description of frequencies
in use within the local region. Furthermore, we assume that the infrastructure beacon is followed by an interval in which CRs request an available time-frequency slot
and associate a net name to that slot, as well as to their location and transmit power,
followed by a response from the infrastructure recommending which frequency to
use and when to check back. Subsequent requests by other net members can be
directed to the proper time-frequency slot, and the geographic distribution of net
members can be tracked, allowing the infrastructure to assess interference profiles.
Unaided Rendezvous
Defense systems are rarely able to assume support infrastructure. Similarly, early
deployments of commercial CR equipment will not be able to assume infrastructure. Consequently, it is important to have an unaided method for rendezvous.
Several methods exist for systems to find each other.
The problem is somewhat like two men searching for each other when they
are a mile apart on a moonless dark night out in the desert. Each has a flashlight
the other may see, but only when it is pointed in the right direction. They can
finally signal each other, but only when each has noticed the other’s flashlight, so
each must look in the proper direction at the proper time.
In the case of CRs trying to find each other, one must transmit and the other
must receive in the proper frequency “hole” at the proper time. Several methods
are feasible. All involve one node transmitting “probe” signals (a uniquely distinguishing waveform that can be readily correlated) until receiving a response from
a net member. In all cases, the problem is that the frequencies the transmitter sees
as being “usable holes” are different from those the receiver sees. Thus, neither
can be stationary and just sit on one frequency hoping the other will find it:
Case 1: Node A transmits probes in randomly selected frequency holes, while
node B listens to random frequencies considered to be unoccupied. Preferably,
154
Cognitive Radio: The Technologies Required
node B listens about three times as long as a probe pulse. The search time can
be dramatically reduced if node B is able to listen to many simultaneous frequencies (as it might with an OFDM receiver). Node B responds with a probe
response when node A is detected.
Case 2: Node A selects the five largest unoccupied frequency blocks, and rotates
probe pulses in each block. Similarly, node B scans all of the unoccupied frequency blocks that it perceives, prioritized by the size of the unoccupied frequency block, and with a longer dwell time.
Both of these cases are similar, but minor differences in first acquisition and
robustness may be significant.
After node B hears a probe and responds with a probe acknowledge, the
nodes will exchange each node’s perception of locally available frequencies.
Each node will “logical OR” active frequencies, leaving a frequency list that
neither node believes is in use. Then node A will propose which frequency
block to use for traffic for the next time slot, at what data rate, and in what
waveform.
From this point forward, the nodes will have sufficient connectivity to track
changes in spectral activity and to stay synchronized for which frequencies to
use next.
4.6.4 CR Authentication Applications
A CR can learn the identity of its user(s). Authentication applications can prevent
unauthorized people from using the CR or the network functions available to the
CR. This enhanced security may be exploited by the military for classified communications or by commercial vendors for fraud prevention.
Because many radios are usually used for voice communications, a microphone often exists in the system. The captured signal is encoded with a VoCoder
(voice coder) and transmitted. The source radio can authenticate the user (from a
copy of the data) and add the known identity to the data stream. At the destination
end, decoded voice can be analyzed for the purposes of authentication, and the
result may be correlated with the sent identity.
Other sensors may be added to a CR for the purposes of user authentication.
A fingerprint scanner in a push-to-talk (PTT) button is not intrusive. Automatic
fingerprint correlation software techniques are available and scalable in terms
of reliability versus processing load required. Additionally, cell phones have
been equipped with digital cameras. This fingerprint sensor coupled with facial
155
Chapter 4
recognition software may be used to authenticate a user. Again, the reliability is
scalable with processor demands.
Figure 4.12 shows a context diagram for a CR’s authentication application.
The detailed learning associated with adding a user to an access list through a
“third-party” introduction is not shown. The certificate authority could be the third
party, or another authorized user could add a new user with some set of authority.
The CR will learn and adapt to the new set of users and to the changing biometric
measures of a user. For example, if a user gets a cold, his voice may change, but
the correlation between the “new voice” and the fingerprint scanner is still strong,
and the CR may choose to temporarily update the voice print template.
User
Microphone
Sensors
Specialized
Biometric Sensors
CR
Authentication
Applications
Handshaking
with Remote CR
Access to
CR Functions
Figure 4.12: Authentication context diagram. A CR application that learns the identity of its
user(s) and has an Access List Security Protocol can minimize fraud and maximize secure
communications.
4.7 Timeline for CRs
CR development will be a spiral effort. The when, where, how, and who of this
development are discussed below:
When: Currently, several CR initiatives are under way. Progress is evident every
day, and more and more sophisticated demonstrations are imminent. Some of
these demonstrations will include better policy conformance characteristics.
Where: The FCC and NTIA are currently discussing a test band for CR. They are
suggesting a 10 MHz experimental chunk of spectrum to allow developers to
experiment. This promising development must be exploited. As policy is written for this band, it can be deployed and policy violations may be assessed.
Who: Vendors, regulators, service providers, and users are highly interested in
CR systems. A great deal of discussion on exactly what that means has
156
Cognitive Radio: The Technologies Required
already taken place and continues. Academic researchers using COTS SDR
demonstrations and government-sponsored demonstrations using customdeveloped SDRs will reduce CR to practice, for at least the first systems.
Where the technology will progress is difficult to predict. As an example, the
following organizations are working with CR: General Dynamics; Shared
Spectrum; Raytheon; Lockheed Martin; Bolt, Beranek, & Newman; Rockwell
Collins; Harris; Virginia Tech; and Northeastern University (and the list has
doubled during the preparation of this book).
How: The easiest experiments utilize COTS SDR hardware and execute cognitive
applications as demonstrations. Custom-developed hardware is more expensive, but is better tailored to show the benefits of CR. Spectrum awareness
using a spectrum sensor has the best information to exploit unoccupied spectrum. This custom capability also has computational resources sized to execute a policy constraint engine.
4.7.1 Decisions, Directions, and Standards
Numerous organizations and standards bodies are working in the area of CR. The
SDR Forum, the Institute of Electrical and Electronics Engineers (IEEE), the
Federal Communications Commission (FCC), the National Telecommunications
and Information Administration (NTIA), and the International Telecommunication Union (ITU) all have interests in this area.
4.7.2 Manufacture of New Products
Many products have new sensors and actuators in them. Cellular telephone handsets are a high-volume, highly competitive product area, from which innovation is
driven by these characteristics. In fact, a large fraction of new cell phones have an
integrated digital camera. These are manually operated today, but CR applications
may take advantage of this sensor for more “cognitive” operation modes.
Chemical sensors have been integrated to some cell phone models. The purpose of the chemical sensors is to report on important “problems” such as “bad
breath” and “blood alcohol level.” This is a manually operated device, but future
applications may be more autonomous.
Among the new applications in cellular telephones is a Bluetooth-like waveform to introduce single people. “Flirting radios” may subsequently need to have
AI technology added to filter out the “undesirable” introductions.
Much to the dismay of teenagers and employees everywhere, phone tracking
applications are now available. Although these capabilities are used “manually”
157
Chapter 4
now, learning algorithms can be applied to the interface to create a new “filtered”
report stream. There is serious interest in tracking other sets of people, such as
first responders, delivery people, service people, or doctors.
4.8 Summary and Conclusions
Radio evolution has taken the path toward more digital realizations and more software capabilities. The original introduction of software made possible software
capable radios that communicate and process signals digitally. In the pursuit of
flexibility, software programmable radios, and the even more flexible SDR, have
become the standard in the military arena and are starting to gain favor in the
commercial world, as explained in Section 4.2. We are now seeing the emergence
of aware radios, adaptive radios, and finally CRs, and we have traced the continuum among these various degrees of capability, as well as providing a few examples of each. Sections 4.3 and 4.4 explored the properties and capabilities of each
of these classes of radios.
Section 4.5 outlined the enabling technologies for CRs. Numerous technologies
have matured to the point where CR applications are possible and even attractive.
The ability to geolocate a system, sense the spectrum, know the time and date,
sense biometric characteristics of people, access new software capabilities, and
determine new regulatory environments are all working together to enable CR.
Geolocation through the use of GPS or other methods is now available.
Chapter 8 discusses how a radio can know where it is located. This enabling technology allows a radio to make spatially variant decisions, which may be applied to
the selection of policy or networking functions.
Sensing of the local RF environment is available. This information may be
used to mitigate a deeply fading channel or may be used to access locally unoccupied spectrum. Non-interference is particularly important, and protocols for sensing, deciding, and accessing spectrum are being designed, developed, and
demonstrated today.
Increased robustness in biometric sensor technology provides a whole new
dimension to CR applications. The most likely initial use of this technology is in
user authentication applications, such as the purchasing of services and products.
Knowledge of time has been available in many forms, but integration into a
broader range of full-function capabilities will enable all new applications. Stable
time knowledge enables a CR to plan and execute with more precision. Using this
capability for non-infrastructure-based geolocation, dynamic spectrum access, or
AI planning is envisioned for near-term CR functions.
158
Cognitive Radio: The Technologies Required
A smart agent model of CRs is attractive. An agent is an entity that perceives
and acts on behalf of another. This is where CRs are going. Smart agent models of
the world enable a radio to provide services for its user. Improved performance or
new capabilities may be provided. As the CR’s smart agent model of the world
becomes more sophisticated and realistic, situational awareness will increase. As
the models improve, the ability of the CR to act effectively over a broader range
of user services will improve.
Maybe the most important software technology is a policy engine that enables
a CR to know how to behave where it is right now, given the priorities of current
circumstances. AI applications at a very high level, networking services in the
middle levels, and signal processing primitives at a very low level are all available
for a CR developer to utilize in creating new capabilities. Finally, middleware
technology enables greater software reuse, which makes CR development
economical.
Modern regulatory philosophy is starting to allow CR to deploy new services
and capabilities. As the trend continues, there are economic motivations for
deploying CR systems.
Section 4.6 covered research in CR technologies, and presented three significant classes of CR applications. Geolocation-enabled applications and authentication applications were discussed in some detail. The most promising CR
application is dynamic spectrum access. Suggestions for using OFDM waveforms
along with dynamic bit loading are included in this chapter. Solutions to the rendezvous problem are suggested, and the hidden node problem is described.
Section 4.7 covered the timeline in which these technologies will roll out and
be integrated into radio equipment and products. Many of the technologies
required to provide some of the useful and economically important CR functions
already exist, so some of these features should begin to appear within the timeline
of the next development cycle.
The bottom line is that the enabling technology for CR applications is available. There is interest in integrating the technologies to build cognitive applications. Finally, the emergence of CRs and their cognitive applications is imminent.
References
[1] Federal Communications Commission, Notice of Proposed Rule Making, August
12, 2000.
[2] http://www.fas.org/man/dod-101/sys/land/sincgars.htm
[3] http://www.fas.org/man/dod-101/sys/land/.htm
159
Chapter 4
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[26]
[27]
[28]
[29]
[30]
[31]
[32]
[33]
[34]
http://www.fas.org/spp/military/program/com/an-wsc-3.htm
http://www.fas.org/man/dod-101/sys/ac/equip/an-arc-164.htm
http://www.fas.org/man/dod-101/sys/ac/equip/an-arc-220.htm
http://www.harris.com
http://www.columbiaelectronics.com/motorola_LST_5b___LST_bc.htm
http://army-technology.com
http://www2.thalescomminc.com
http://www.fas.org/man/dod-101/sys/ac/equip/csel.htm
www.jcs.mil/j6/cceb/jtidsmidswgnotebookjune2005.pdf
http://www.fas.org/man/dod-101/sys/ac/equip/an-arc-210.htm
http://www.afcea.org/signal
http://www.fas.org/man/dod-101/sys/land/eplrs.htm
http://www.nj7p.org/history/portable/html
http://raytheon.com/products/jtt_cibs/
http://www.acd.itt.com/case4.htm
http://jtrs.army.mil/
http://jtrs.army.mil/
http://www2.thalescomminc.com
www.comsoc.org/tech_focus/pdfs/rf/06.pdf
http://www.gdc4s.com/
http://jtrs.army.mil/
http://www.gnu.org/software/gnuradio
http://vanu.com/technology/softwareradio.html
J. Polson, “Cognitive Radio Applications in Software Defined Radio,” in Software
Defined Radio Forum Technical Conference and Product Exposition, Phoenix, AZ,
November 15–18, 2004.
H. Chen, S. Kumar and C.-C. Jay Kuo, “QoS-Aware Radio Resource Management
Scheme for CDMA Cellular Networks Based on Dynamic Interference Guard
Margin (IGM),” Computer Networks, Vol. 46, 2004, pp. 867–879.
B. Bougard, S. Pollin, G. Lenoir, L. Van der Perre, F. Catthoor and W. Dehaene,
“Energy-Aware Radio Link Control for OFDM-Based WLAN.” Available at
http://www.homes.esat.kuleuven.be/bbougard/Papers/sips04-1.pdf.
http://www.darpa.mil
P.H. Dana, The Geographer’s Craft Project, Department of Geography, University
of Colorado, Boulder, 1996.
E. Azzouz and A. Nandi, Automatic Modulation Recognition of Communication
Signals, Springer, New York, 1996.
S. Harris, CISSP All-in-One Exam Guide, 2nd Edition, McGraw-Hill, New York,
2002.
http://www.bbn.com
160
Cognitive Radio: The Technologies Required
[35] S.J. Russell and P. Norvig, Artificial Intelligence: A Modern Approach, 2nd Edition,
Pearson Education, London, 2003.
[36] Federal Communications Commission. Spectrum Policy Task Force Report. ET
Docket No. 02-135, November 2002.
[37] Hughes-Hartogs, “Ensemble Modem Structure for Imperfect Transmission Media,”
United States Patent 4,679,227, July 7, 1987.
[38] P. Chow, J. Cioffi and J. Bingham, “A Practical Discrete Multitone Transceiver
Loading Algorithm for Data Transmission Over Spectrally Shaped Channels,”
Transactions on Communications, Vol. 43, February/March/April (2/3/4), 1995,
773–775.
[39] B. Krongold, K. Ramchandran and D. Jones, Computationally Efficient Optimal
Power Allocation Algorithms for Multicarrier Communications Systems, University
of Illinois, Urbana–Champaign.
161
This page intentionally left blank
CHAPTER 5
Spectrum Awareness
Preston Marshall
Defense Advanced Research Projects Agency
US Department of Defense Arlington, VA, USA
Spectrum is the “lifeblood” of RF communications.
5.1 Introduction
The wireless designer’s adaptation of the classic New England weather observation
could be “Everyone complains about spectrum availability (or at least the lack of it),
but no one does anything about it!” Cognitive radio, however, offers the opportunity
to do something about it. Spectrum aware radios offer the opportunity to fundamentally change how we manage interference, and thus transit the allocation and utilization of spectrum from a command and control structure—dominated by decade-long
planning cycles, assumptions of exclusive use, conservative worst-case analysis,
and litigious regulatory proceedings—to one that is embedded within the radios,
each of which individually and collectively, implicitly or explicitly, cooperates to
optimize the ability of the spectrum to meet the needs of all the using devices. As
this chapter looks at this opportunity, it investigates solutions that range from
local brokers that “deal out” spectrum, to totally autonomous systems that operate
completely independently of any other structures. In its ultimate incarnation, it is
possible to actually use spectrum awareness and adaptation to relax the hardware
physical (PHY) layer performance requirements by avoiding particularly stressing
spectrum situations. As such, a cognitive radio could ultimately be of lower cost
than a less intelligent, and more performance stressed, conventional one.
Note: Approved for public release, distribution unlimited.
163
Chapter 5
5.2 The Interference Avoidance Problem
Before discussing a cognitive spectrum process, let us consider the classical spectrum management and assignment case. Once radios moved beyond spark gap
techniques (the original impulsive ultra-wideband (UWB) radio), use of the spectrum
has been deconflicted to avoid interference. Spectrum and frequency managers
assign discrete frequencies to individual radios or networks and attempt to ensure
that the emissions from one do not adversely impact others. A not insignificant
legal (and seemingly smaller technical) community has grown up around this simple principle. Such planners are inherently disadvantaged by a number of factors.
For one, they have to assume that:
●
●
Interfering signals will propagate to the maximum possible range.
Desired signals will be received without unacceptable link margin degradation.
In practice, this means that interference analysis is often driven by two
unlikely conditions: maximal propagation of interfering signals and minimal
propagation of the desired signal. Simplistically, if propagation was a simple
R2 condition (from a tower to a close-by remote), and the receiver needed 12 dB
signal-to-noise, then we could consider that the interference would extend for a
distance of four times the maximum range of the link, at which point the received
signal power would be equal to the noise power.1 Thus, the interference would
be an area 16 times the usable coverage area, as shown in case 1 in Figure 5.1.
Operation close to the ground would have an R3.8 propagation, and that ratio
Figure 5.1: Practical interference
margins. Several cases of
conservative assumptions
about propagation to intended
and unintended receivers are
illustrated.
Intended
Communications
Range
Case 1 Advantaged
Victim
Case 2 Equally
Disadvantaged
Victim
Case 3
Disadvantaged
Victim, with Fade
Margin Allowance
1
For an explanation of propagation energy loss, refer to the Appendix at the end of this chapter.
164
Spectrum Awareness
would be reduced to about four, as shown in case 2, which is slightly less
stressing. However, when we consider that one of the receivers or transmitters
may be at an advantaged position and one at a disadvantaged position (and thus
has losses approaching R3.8), and when we must include a multipath fading
allowance on the desired link, the relative gap opens up immensely. This is shown
in case 3. If we assume a worst-case fade margin (a possibility of 20 dB loss to the
disadvantaged intended receiver operating in an R3.8 regime, while having other
unintended receivers operating at R2 in the same region), it increases the area of
interference to more than 150 times the area of reliable communications. Figure
5.1 illustrates the impact of ever-increasing conservative assumptions regarding
signal propagation, and imputed spectral usage for conservative spectrum planning situations.
When we add mobility to this analysis, the situation gets even more restrictive.
In this case, the conventional frequency manager must deconflict the entire
range of motion of the emitters and receivers. All this greatly consumes effective
spectrum, increasing the already low (only several percent average utilization) values of the pessimistic and static case, to add the considerations of pessimistic
assumptions about possible mobility.
It is important to recognize that this pessimism is not inherent in the operation
of the radio links; statistically, it is important because it represents a set of cases
wherein many radio systems would have no ability to operate, because without
adaptation, even if the radio recognizes that an interference condition exists, it
cannot implement and coordinate a strategy for migrating to a clear channel. In
this chapter, we consider spectrum strategies that use awareness to locate spectrum holes, often the result of the essential conservative nature of the planning
process. However, an equally important rationale for the inclusion of awareness
in real systems is the ability to locally resolve interference by using the same
behaviors to locate new, and unblocked, spectrum. This feature of interference
adaptive radios offers all users of the spectrum the ability to migrate from the
currently conservative assumptions that underlie spectrum planning.
5.3 Cognitive Radio Role
We postulate that cognitive radio offers the ability to manage this situation more
effectively by utilizing the ability to sense the actual propagation conditions that
occur, and to adjust the radio dynamically to best fit these conditions. To do this,
we distinguish between two radio operating objectives. In the first, the radio
attempts to minimize its own spectral “footprint,” consistent with the environment
165
Chapter 5
and needs of the networks it supports. In the second, it adapts itself to fit within
whatever spectrum is available, based on local spectral analysis. When we put
these two together, we can conceive of a radio that can find holes, and then morph
its emissions to fit within one or more of these holes. Such radios could offer
radio services without any explicit assignment of spectrum, and still be capable of
providing high-confidence services.
5.4 Spectral Footprint Minimization
There are two sides to the cognitive radio problem. The first is fitting into the
spectral footprint of other radios. The second, and more subtle, is to minimize the
radio’s own footprint, as discussed here.
For years, modulation designers have defined spectrum efficiency as the number of bits per hertz (Hz) of bandwidth, and have often used this metric as a scalar
measure of the best modulation. The assumption has been that the design that utilized the least spectrum was intrinsically less consumptive of shared spectrum
resources. This proposition is worth examining. The Shannon bound argues that
essentially infinite bits per Hz can be achieved, but it can be accomplished only by
increasing the energy per bit (Eb) exponentially, and thus the radiated spectral
energy increases at the third power of the spectral information density.2 We broaden
our view of spectrum impact to include not only the amount of spectrum used, but
also the area over which it propagates.3
One problem involves how to define spectral efficiency metrics. Classically,
digital radio engineers have attempted to minimize the spectrum used by signals
through maximization of bits per Hz. This is a simple and readily measured metric, but is it right to apply this to a new generation of radios? The Shannon bound
shows a basic relationship between energy and the maximum possible bits per Hz,
as shown in Figure 5.2.
An immediate observation is that although the bits per Hz increase linearly,
the energy required per bit goes up exponentially, as energy increases an average
of 2 dB per bit per Hz over the entire range of spectrum efficiencies. We can see
that the proportionate cost in energy of going from one bit to two bits per Hz is
essentially the same as that required to go from six to seven. Essentially arbitrary
bits per Hz are possible if the channel is sufficiently stable and power is available.
But this is tough on battery-powered devices!
2
The Shannon bound (or limit) is the best that can be done at this time.
It is worth noting that volume would be a more generalized measure, but most spectrum usage is
on the surface of Earth, so we will limit our consideration to two, rather than three, dimensions.
3
166
Spectrum Awareness
Figure 5.2: Shannon EB for various
spectral efficiency values.
25
Eb/No (dB)
20
15
10
5
0
5
0
2
4
6
8
10
12
8
10
Bits per HZ
1.2
Bits per Unit Area
Figure 5.3: Impact of bits per Hz on
spectrum footprint.
1
R4
0.8
R3
0.6
R2
0.4
0.2
0
2
4
6
Bits per HZ
It is equally tough on the other users of the spectrum, in the form of suppressing adjacent channel interference, as well as large dynamic range in the analog
functions of the receiver. In order to increase the bits per Hz, the radio must now
transmit both slightly more bits, and vastly more Eb, in order to meet the Eb/No
requirements of the receiver, where No is noise power.
Using Shannon’s limit, we can compute the increase in spectral energy
required to increase spectrum efficiency from one-half to eight bits per Hz. If
simple spectrum were the measure, such a radio strategy would be effective.
However, for most frequencies, spectrum is a valuable asset over a given area.
Therefore, bits per unit area is an appropriate measure for how effectively we use
spectrum. Extending this bound to consider propagation is important if we want to
understand how this denies spectrum usage to the radios. Ignoring multipath and
absorption, propagation between terrestrial antennas can be simply modeled
as R, with varying from 2 to 4, depending on antenna height and frequency.
Applying Shannon’s limit to this simplistic propagation yields the graph in
Figure 5.3. In this case, we compute the change in effective interference region
for each point in the Shannon curve, and for cardinal values of the propagation
167
Chapter 5
constant. Clearly, simplistic spectrum strategies that only look at bits per Hz are
not effective, and in fact are quite counterproductive for all the spectrum users as
a whole. The better the propagation, the more ineffective increasing energy is as a
strategy for spectrum efficiency.
For most propagation conditions, it is clear that increasing the spectral
efficiency of one radio will disproportionately reduce the ability of the spectrum
to support other users. The value of generally increases with frequency (for
propagation in very high frequencies (VHF) and above), which implies that more
sophisticated strategies are needed. Increasing modulation constellation depth is a
poor solution for the radio because it greatly increases the energy it needs, as well
as a poor solution for the other radios sharing the spectrum because it essentially
raises the noise floor throughout a greatly increased region, or precludes operation
at rapidly increasing radii from the transmitter.
This can lead us to a measure that recognizes this trade-off. In this case, bits
per Hz per area reflects that the critical issue in spectrum optimization is not
spectrum use; rather, it is spectrum reuse, and we want to measure and optimize
not only how the radios themselves perform, but how the radios allow other
radios to share the spectrum in a close to globally optimal manner.
The preceding discussion assumes that the power is perfectly matched to the
channel; in practice, such an alignment is impossible to achieve or maintain over
any degree of time-varying channel, so the real result will fall below these values.
In the extreme case of no power management, the results are insensitive to many
of these considerations.
5.5 Creating Spectrum Awareness
One of the key considerations for whether a radio is cognitive is its ability to create awareness of its environment. Certainly, the key environment of a radio, or
network, is the spectrum in which it operates.
5.5.1 Spectrum Usage Reporting
The simplest implementation of a cognitive spectrum architecture could be based
on principles already established for the upper layers. Each radio would coordinate the use of spectrum via spectrum usage (spectral power density, directionality, location) reporting and distribution. Reporting to local neighbors or to a local
spectrum management infrastructure would have to be provided not only by the
transmitters, but by the receivers as well. Receiver reporting would enable the
168
Spectrum Awareness
sharing scheme to be more aggressive because the system would have to protect
only locations in which a possible interference could occur. Without receiver
reporting, the interference analysis would have to assume receivers throughout the
entire radiated pattern. A collection of all such reports (and their associated
locations) would, in theory, provide each member of the network with an exact
understanding of the interference it would face and the interference it would cause
if it radiated any specific emission. Such architectures could resemble the
command and control model of spectrum, if the coordination was provided to
a single authority, or “band manager,” that would control the dynamic assignment
of spectrum.
Alternatively, the control could be more peer to peer, with individual radios
computing the effect on other radios as well as on themselves. This approach differs from the current mechanisms not so much by its intrinsic character, but by its
automation, time cycle, and “localness.” It meets the definition of cognitive in that
it makes decisions based on a level of awareness provided by other users of the
spectrum. In this model, individual radios are blind. Such a scheme has the advantages cited above if (and only if) the structure can meet the following conditions:
●
●
●
Fellow radios are assumed to be trustworthy.
The general characteristics of the propagation environment are well understood,
or the bounds on the best possible performance are well understood.
Reporting requirements are imposed on 100 percent of the emitters (and
optimally, the receivers).
The process is certainly appropriate for application in bands in which the
usage is homogeneous and controlled by a single entity (or possibly a set of
cooperating entities), and in which the antennas in at least one end of the link are
generally advantaged, in order to have the bounding equations somewhat
correspond to likely propagation.
5.5.2 Spectrum Sensing
There are two current technologies for spectrum sensing: (1) Conventionally,
spectrum analyzers have been rapidly tuned through a band, with the scan frequency adjusted to provide adequate “dwell” on each frequency bin, but scanning
through all of the channels rapidly enough to be “real time” in terms of the signals
of interest. (2) More recently, analog-to-digital conversion (ADC) has made the
use of the fast Fourier transform (FFT) practical. This sensing technology has the
169
Chapter 5
advantage that it provides instantaneous analysis of all the frequencies within the
FFT window. Both technologies have advantages, depending on the application,
cost, energy available, and type of resident signal. Table 5.1 illustrates some of the
qualitative advantages of each.
Table 5.1: Comparison of spectrum sensing approaches.
Technology
Spectrum analyzer
Wideband FFT
Complexity Lower complexity if it can
share receiver components
Short signal Low dwell time on each
detection
frequency can fail to see short,
pulsed signals. Typical dwell is
only microseconds to
milliseconds, and below 1%
Bandwidth Can scan large ranges of signal
Speed
Slower, based on dwell time
on each channel. May be
difficult to interleave “listen
through” without penalty to
the node performance
Power
Mostly classical analog
components, potentially
shared with the mission
receiver
Adds requirement for wideband
ADC and extensive DSP
Higher probability of detecting
short signals, based on duty
cycle of the sensing, which
could reach 50%
Limitations in ADC constrain
instantaneous bandwidth
With appropriate filters, sample
sub-Nyquist to minimize time
delay and interleaving time.
These short intervals can be
compatible with MAC layer
timing
Digital processing adds to the
inherent analog energy usage
MAC: medium access control; DSP: digital signal processing.
5.5.3 Potential Interference Analysis
We can consider interference in two categories. The first category is direct interference with ongoing communications of the primary user, resulting in degraded
communication functionality. The second category is interference to the channel
when it is not in use, causing the primary user to think there is something wrong
with the channel or with the equipment.
The first category involves direct interference with the quality of the signal
at one or more layers of the receiver, as shown in Table 5.2. In conventional
170
Spectrum Awareness
Table 5.2: Interference effects on digital processing layers.
Layer
Impact
Mitigation
PHY
Creation of higher
uncorrected BER
MAC
Complete and uncorrectable
blockage of a packet
NETWORK
Complete and uncorrectable
blockage of a packet
TRANSPORT
Network fails to route
packet
Allow higher layers to resolve a
short interval of interference
Adjust coding dynamically
MAC layer acknowledgment
retransmits. Blocked RTS are
automatically retried
Protocol operates effectively
with missing data. For example,
VoIP can suffer some loss as
long as it is not correlated
Transport layer recognizes
missing sequence and requests
retransmission
BER: bit error rate; VoIP: Voice over Internet Protocol.
spectrum practice, interference has generally been very quantitatively defined as
the ratio of signal energy to the signal plus interference. Although this definition
is scientific, the regulatory community has extensive filings from spectrum users
demonstrating either extensive additional energy, or minimal energy, often in the
same situation. Unprotected receivers have a cascading effect. Very small errors in
the lower layers can have disproportionate impacts on the upper layers. For example, a few short bursts of noise may not impact a voice channel, but if it introduced a 103 error rate in a digital packet communications link, it would
essentially deny functionality. Similarly, a packet loss rate of 101 might have
minimal impact on the network layer, but could, if timed badly, essentially shut
down a transport layer, such as Transmission Control Protocol (TCP). In contrast,
if cognitive radios are the norm, these same layers can assist by reducing the
unavoidable temporary interference that may occur at lower layers. Table 5.2
shows examples of how the various layers can magnify interference events or can
mitigate the effects. The eventual adoption of interference-tolerant systems creates
an environment in which less conservative practices can be applied, because in the
rare cases when interference is caused, the victim radio can address this locally
without losing its effectiveness.
A second potential source of impact is more subtle. This is the effect of the
use of given spectrum on other (and assumed more primary) users. In this case,
171
Chapter 5
we contemplate impacts that are short of specific denial of communications, and
that may not be measurable at the product of the communications system. For
example, if a push-to-talk (PTT) frequency is not in use by any transmitter, and
we “borrow” it and thereby cause all of the voice receivers to emit a horrible
shrieking sound as the squelch is broken, we have caused interference to another
communications network. This problem is unique in that it requires us to consider
that the absence of any signal constitutes communications in and of itself, and that
any use of the frequency is therefore interfering with the communications. This
implicit interference is one that cannot be dealt with in the same way as direct
interference.
One way to accommodate the non-quantitative nature of interference is to
consider the evolution of bands shared by cognitive radios. In the initial deployment of such systems, existing legacy users will be the dominant users of the
bands. These radios were not designed to mitigate the very unnatural effects of
cognitive radio sharing, so cognitive radios may require care and extensive measures to ensure minimal impact. (An exception may be military radios, which are
designed to have extensive interference tolerance to avoid electronic attack, such
as jamming.) As cognitive radios become the majority, a principle such as “5 mile
an hour bumpers” could be adopted. In this principle, radios would have some
assumed ability to mitigate the effects of interference by using adaptation at all
layers so that they could be tolerant of the less than perfect environment that
extensive use of cognitive radios may create.
This posits the following cascading set of questions:
●
●
●
If (using 802.11 as an example) I interfere with a single request-to-send (RTS)
message, but you get the next one through, have I interfered with you?
If I added Gaussian noise, but it was within your power management capability,
have I interfered with you?
If I added some Gaussian noise to your channel, but it was well within
your margin and was within the range of your error correction, have I impacted
you?
Spectrum owners that paid billions for spectrum would certainly say yes to
all of these because it impacted the capability that they could achieve with their
investment, but a community of band sharers may consider it just a cost of being
in a shared band. Measuring spectrum is fundamentally a technical process, but
assessing the spectrum is a much more complex perceptual, legal, and relativistic
process.
172
Spectrum Awareness
One of the best educations that can be achieved in the analysis of interference
is provided by reading the voluminous filings before national regulatory agencies.
Well-qualified, and (we should assume) honest engineers can take the same set of
facts, apply well-known engineering principles, and can reach diametrically
different conclusions on the simple question of whether a certain emission will
cause interference to another!
5.5.4 Link Rendezvous
When an adaptive radio is first turned on, it faces a unique burden: how to find
the other radios with which it is intended to communicate. Conventional radios
generally have some prior knowledge of frequency, or the frequency is provided
by the operator. In the case of a cognitive radio, no one knows what the best
frequency is until the environment is sampled. This topic is further discussed in
Chapter 8.
5.5.5 Distributed Sensing and Operation
Distributed operations occur when the spectrum awareness problem leaves the
domain of signal processing and enters the realm of cognitive processes. There
are inherent limitations in the ability of any radio to fully understand its environment. One of the fundamental issues with the idea of radios creating spectral
awareness is that individual radios, typically close to the ground, are subject to a
wide range of propagation effects, all of which appear to be conspiring to maximally disrupt the radio’s ability to create an understanding of the spectral environment. The transition from direct to diffracted communications is a major effect
driving uncertainty. Direct propagation follows something like the PHY layers
R2 rule, whereas diffracted (non-direct) propagation attenuates at the rate of
approximately the fourth power of distance. The inability of the radio to know
which of these propagation conditions is present affects its ability to sense, determine interference, and assess the consequences of its own communications. The
propagation of signals involves five parties, including a victim transmitter, victim
receiver(s) (of unknown location or even existence), a cognitive transmitter, a cognitive receiver, and a spectrum sensor (often assumed to be, but not necessarily
located at the Defense Advanced Research Projects Agency (DARPA) NeXt
Generation (XG) transmitter location). A matrix of possible propagation
conditions between these nodes can be developed, with the possible values of
the matrix characterizing the nature of the propagation between each (R2, R4, R4
plus 20 dB multipath fade, etc.).
173
Chapter 5
At the onset, the cognitive radio has no awareness of how to fill in this matrix,
and so must assume worst-case conditions on each link, which might be:
●
●
●
R2 propagation from the cognitive transmitter to the victim receiver;
R4 plus 20 dB multipath fade from the victim transmitter to the spectrum
sensor;
R4 plus 20 dB multipath fade from the cognitive transmitter to the cognitive
receiver.
Removal of uncertainty can greatly reduce the power that the cognitive
transmitter can emit with certainty of non-interference to the victim receiver. In
this case, the assumption of multipath cannot be eliminated, but multiple sensors
measuring the same emission can greatly reduce the potential range of the
assumptions that the interference analysis must address. For that reason,
collective sensing, particularly in a mixed direct and diffracted multipath region,
may be essential in allowing cognitive radios to emit appropriate energy levels.
This implies that additional algorithms will be needed to fuse this sensor data
and establish acceptable confidence bounds for the inter-radio propagation
assumptions.
5.6 Channel Awareness and Multiple Signals in Space
The implicit assumption of most spectrum-sharing approaches is that a given
frequency can be used by only one spectrally efficient signal (narrow bandwidth)
at a time. The concept of multiple input, multiple output (MIMO) signaling, first
described as BLAST, has been subsequently generalized and is beginning the
process of commercial exploitation. In MIMO, multiple signal paths create separable orthogonalized channels between multiple transmit and multiple receive
antennas. Although this process at first appears to violate Shannon’s bound, in fact
it is only sidestepping Shannon because in theory, each of the reflected
paths is an independent channel. Essentially, this technique turns multipath into
multiple channels [6].
Receivers nearly always benefit from knowledge of the channel. Active equalization, RAKE filters and other techniques exploit this awareness to improve performance at the link layer of the radio. These techniques are invisible to the upper
layers. MIMO provides an important distinction from the traditional modem
design. In the case of MIMO, we use the channel awareness much more architecturally, in that we design the links, and potentially the topology of the network, to
174
Spectrum Awareness
be dependent on this multipath, rather than just tolerant of it. A very large body of
published literature describe MIMO more fully [1–6], and therefore this discussion is not elaborated here.
Certainly MIMO can be implemented without cognitive features; however,
these implementations are incapable of providing an assured level of performance
because the repairability of the channel cannot be known in advance. No amount
of processing can create 10 independent channels in environments that have no
reflective component. However, when we combine MIMO with other spectrally
adaptive techniques, it is now possible to create assured services by balancing the
techniques of spectrally adaptive and spatially adaptive techniques.
Consider the reasoning at the radio and network levels. The radio operates
best (least energy investment per bit) in a clear channel, without the complexity
and induced noise of MIMO techniques. A “greedy” radio would never choose to
be a MIMO radio. It is only when the needs of individual radios are balanced with
the spectrum usage of other devices that the radio benefits from MIMO. Chapter 11
further discusses this from a network perspective, but even from a radio perspective there are advantages to MIMO when the spectrum available to the radio is
capped. As an example, MIMO is attractive in unlicensed bands, such as used by
802.11, because growth in capability must be achieved within a fixed spectral
footprint.
Spectrum awareness therefore has two components. If adaptive spectrum techniques are permitted, then the spectrum awareness function provides an assessment of the relative scarcity of spectrum resources. This relative availability
measure provides an implicit or explicit metric regarding the cost that the radio
should be willing to pay to achieve the spectrum usage reductions from MIMO.
The other aspect of spectrum awareness is the measure of path differentiation
available from the apertures of the radio. This measure is more specific to the
MIMO algorithm, but in general can yield a metric reflecting the costs and benefits of various degrees of MIMO usage.
Cognitive technology is critical if the radio is to apply techniques more complex than simply utilizing all of its resources to the maximum extent in MIMO
processing. In theory, MIMO provides a linear growth in throughput as the number of antennas are increased. In practice, the channels between the elements may
be less than ideally separated. Figure 5.4 illustrates the assessment of throughput
versus channel usage for a spectrum- and power-constrained radio. The upper
curve represents the theoretical bound of “perfect MIMO,” and the lower curve
represents a situational assessment that a radio might make, given its ability to
learn the degree of separation. Note that the marginal benefit of channels
175
Chapter 5
Figure 5.4: MIMO capacity versus
transmit/receive resources.
7
5
Perfect
Situational
3
1
3
5
7
decreases while the resource expenditure increases linearly. A cognitive radio
should use its awareness of channel characteristics and spectrum availability to
assess that it has a number of options beyond “brute force” to employ its resources.
Taking the data from Figure 5.4, we can create a number of different strategies
that a cognitive radio could consider. Each strategy trades some resources (e.g.,
spectrum, throughput, etc.), but all yield more effective use of resources. (Chapter
11 investigates network layer cognitive approaches that could further mitigate
these choices by use of the network selection of alternative topologies.) Table 5.3
shows some obvious choices between MIMO and spectrum usage. An even more
complex, and potentially beneficial, extension of this approach would be to consider varying the Eb in order to also adapt the channel rate. Even this simple
example shows that an adaptive mix of MIMO and spectrum usage is vastly superior to each approach applied separately.
The same data can be plotted to demonstrate the flexibility available to a radio
that can adapt its performance across these two exemplar dimensions, as shown in
Figure 5.5.
5.7 Spectrally Aware Networking
The prior discussion treats the PHY layer as a given that must be constructed in
essentially a fixed fashion. Certainly, this is the current practice, as we direct the
link layer to form certain topologies with explicit, and generally fixed, capacity
between each of the modes. In the ignorance of spectrum conditions, this is as
good an approach as any. However, when we introduce the capability to assess the
costs of each link (e.g., amplifier mode, Eb, etc.), we find that these simplistic
notions of the radio’s role fail to fully exploit the options that we can now create.
176
Spectrum Awareness
Table 5.3: Alternative approaches for mixed spectral and MIMO adaptation.
Channels
used
Resources
1
8
1
2
Pure 8 8
MIMO
6 6 MIMO
2 sets of 3 3
MIMO
2 sets of 4 4
MIMO
4 sets of 2 2
MIMO
8 independent
channels
Throughput
Throughput/
resource
Throughput/
channel
5.04
0.63
5.04
6
6
4.49
5.68
0.75
0.95
4.49
2.84
2
8
7.18
0.90
3.59
4
8
7.80
0.98
1.95
8
8
8.00
1.00
1.00
6.00
Spectral Efficiency
(Throughput per Unit Spectrum)
8 8 MIMO
6 6 MIMO
5.00
2 Sets of 4 4 MIMO
4.00
2 Sets of 3 3 MIMO
3.00
4 Sets of 2 2 MIMO
Cognitive Adaptation
2.00
1.00
8 Independent
0.60
0.65
0.70
0.75
0.80
0.85
0.90
0.95
1.00
1.05
Relative Efficiency
(Throughput to Resources Used)
Figure 5.5: Adaptive trades from MIMO and spectrum usage.
These opportunities are somewhat hidden if we approach networking from the
perspective of the technologies that have dominated wired networking because in
a wired structure the capabilities of each link are intrinsic in the topology. Only in
wireless networks can the network dynamically redefine the capability of each
link based on the needs of the network and the availability of resources.
177
Chapter 5
It is anticipated that early development of adaptive radios will generally result
in “greedy algorithms” that obtain sufficient spectrum for their own needs and are
not cooperative with other spectrum users. It is likely that policy-makers would
force some spectrum churn in order that the spectrum not be filled with first-in
devices that leave no room for new entrants. However, as adaptive devices grow in
use, the opportunity for cooperation is increased, and the ability of the devices to
cooperatively determine networking strategies that are more globally optimized is
presented to the community.
As an example, imagine that we desire to set up a fully adaptive, spectrally
aware cognitive network around the Washington, DC, area. Nodes are distributed
on the ground in density proportional to the population of the region. If we form
the network based on closest neighbor, it would probably tend to route traffic in
and out of the densest areas: into and out of the heart of the city. Although topologically convenient, this network may be unsupportable because it concentrates
both nodes and routed traffic in the same region. A spectrally aware network
might force a quite different organization. This network formation would not
route traffic toward the city; it would route out of the city, around the beltway
(perimeter or ring road for Washington) and back in. This would reduce the peak
spectrum needs in the city and make use of the lower density in the outer regions
for high bandwidth “backbone” communications. This architecture is dependent
on the development of spectrum sensing, automated organization of individual
links, and then the imposition of a global set of objectives for the network that can
be implemented by the distributed nodes.
An unanswered question for research is how similar in technology the network
adaptation is to the link-level adaptation. Because of the policy implications of
spectrum, the link technology has often been driven by the need to comply with
policy, thus favoring declarative technologies, such as reasoning engines. Network
optimization has been a subject of a greater range of technology. The literature
reports a very wide range of techniques, from genetic algorithms, to learning
engines, to declarative solvers as candidate technologies [7–9].
5.8 Overlay and Underlay Techniques
The preceding discussion assumes that communication is organized into discrete
channels that are orthogonal by frequency or time; that is, only a single signal can
occupy a given piece of spectrum at one time. In fact, a number of techniques
allow us to violate this assumption. In the simplest case, the use of one of the
forms of direct sequence spread spectrum (DSSS) offers the ability to transmit
178
Spectrum Awareness
multiple signals on the same frequency. Although a number of signals can share
the same spectrum, the coding added to each signal results in, at best, a zero-sum
benefit. There are significant reasons to use this approach, but they lie in the
enhanced access control, not intrinsic spectrum usage sharing. Furthermore, the
coordination required to meet the stressing requirements of signal power management (to avoid near/far interference) clearly puts many of these technologies in
the medium access control (MAC) layer as much as in the link layer. This chapter
will not further address this technology, although the topic is well developed in
the literature [10–14].
Of more interest to cognitive radio are the techniques that offer the ability to have
signals coexist in space, time, and frequency, without the introduction of coding,
access resolution or mediation, or other mandated control protocols. The US Federal
Communications Commission (FCC) Spectrum Policy Task Force (SPTF) report
[15] describes a general approach, referred to generally as interference temperature.4
The concept of interference temperature created an assurance to band users
that total noise introduced into the band would not exceed a certain spectral noise
temperature, measured as an equivalent radiation temperature. Devices would
have to measure the existing noise temperature and then determine if they could
add more noise (heat in the temperature metaphor) without exceeding the
maximum temperature for the band. This is an important and fundamental change
in the concept of spectrum coexistence in that it has the devices determine their
own effect on the other users of the band and act accordingly, rather than perform
operations to a set limit that was predetermined to have no impact on other users.
Although the algorithms are not necessarily complex, it certainly crosses the cognitive threshold because not only is the radio aware of its environment, it’s behavior is guided by a set of constraints on how its operation is permitted to impact the
environment. Such a regime would inherently resolve the issue of aggregation of
devices that is so much a concern over spectrum sharing.
UWB is the logical extension of direct sequence beyond the point where the
spectral energy density is considered to be a factor in the operation of other UWB
or narrower band receivers. Currently, two general approaches to this technology
are being developed, one using an impulsive signal that is wideband by the nature
of the step function modulation, and one using very wide spreading codes. For
narrowband receivers (seeing only a fraction of the signal energy and for durations significantly longer than the pulse width), both appear as Gaussian noise.
4
Noise temperature is directly related to spectral noise density by the equivalent black body
radiation.
179
Chapter 5
Shannon’s limit promises great benefits from using such wideband signaling, and
in the absence of other users, such approaches offer the potential to remove any
need for spectrum management at all. The inherent constraints of UWB limit its
use to very local signaling, or specialized applications. As such, it appears to be
an extension of networking, and thus exists on the edge of the cognitive networks
that are the focus of this chapter.
5.9 Adaptive Spectrum Implications for Cognitive Radio Hardware
Another aspect of cognitive radio operation will be the minimization of channel
interference due to spurious artifacts inside the radio itself. If all components of
the radio are operating in a linear region, then interference avoidance will be
accomplished by the inherent operation of the radio’s interference avoidance algorithms because channels that contain interference generally are not used, thus
avoiding interference to other users. Although little work has been done on mitigation of nonlinear effects, it is a critical aspect of performance for wideband,
handheld, energy-constrained devices that must operate in the presence of highpower neighbors. In this case, the cognitive process can also be utilized to select
frequencies that have the least intermodulation or internal spur artifacts. Such an
approach offers the ability to maintain connectivity performance even with
reduced analog performance, as long as suitable preselector technology is available. This application of cognitive radio has had relatively little research, and will
not be explicitly addressed further here. A note of interest is that techniques that
were developed to avoid interference to other devices may also have the effect of
providing self-management of internally generated interference as well.
Most of the implications of supporting cognitive radio technology within a
radio have focused on the additional capability required in the radio, and thus the
strong implication that a cognitive radio is inherently more expensive, and thus
limited in application, compared to less capable devices. All other performance
held constant, this implication is certainly true. It is worth investigating the performance drivers for wireless hardware. If we consider a radio, we can partition it
into the analog and digital components. Digital logic is subject to Moore’s Law,
and increasing capability, lower cost, and lower energy are generally available,
just by waiting. For the analog portion, the technology is much more mature, and
it is unlikely that technology by itself will resolve shortfalls with currently available components. Starting with the radio frequency (RF) front-end, several driving
performance characteristics generally drive energy consumption and cost. A very
cursory summary is presented in Table 5.4.
180
Table 5.4: A summary of critical analog performance characteristics.
Meaning
Impact
Constraints
LNA Intercept
Points
(IIP2, IIP3)
SFDR
Measures the linearity of
the receiver’s amplifier
Poor front-end will cause the
receiver to overload,
generating mixing products
Spurs appear as (generally)
tonal signals that can look
like occupied signals, and
can add noise to received
signals
Limits sensitivity to detect
other spectrum users
Increasing IP has severe
penalty in energy use,
thermal load, and cost
Reduction in SFDR requires
significant investments in
receiver organization, packaging,
and components
181
Metric
Noise figure
The maximum dynamic
range with no selfgenerated signals
Measure of selfgenerated noise in the
receiver
LNA: low-noise amplifier; SFDR: spur-free dynamic range; IP: Internet Protocol.
Lower sensitivity leads to
reduction in the power that can
be transmitted for a given
confidence of non-interference
Chapter 5
Although minimal research is currently being performed to validate this
assumption, many of these limitations should be amenable to spectrum awareness
and cognitive reasoning. If successful, this approach could reverse the assumption
that a cognitive radio is inherently more expensive than a less “intelligent” one. If
it can be shown that these adaptation technologies and techniques are effective,
then it is possible that a cognitive radio could enable relaxation of analog performance requirements, and thus make a cognitive radio actually less costly than a
less adaptive one.
5.10 Summary: The Cognitive Radio Toolkit
What sets a cognitive radio apart from a conventional radio may be that, in the
end, it fundamentally changes the way we design radios. Ideally, in fact, we would
not design a cognitive radio at all. We would load it with the range of tools discussed in this and other chapters, and let it design and redesign itself continuously
as both the environment and needs changed. Today, digital communications engineers attempt to system engineer radio operating points in order to draw the best
compromise among competing demands. Because most networks operate within a
relatively static range of characteristics, such operating points are compromises
that are (hopefully) reasonable across the range of operating conditions, but it
would be unreasonable to believe that these points would be optimal in any but a
rare number of situations.
It is highly likely that future networking will continue to move to the adoption
of more and more cognitive technologies. However, if this adaptation is driven by
the needs of fixed networks, it is unlikely to address the unique aspects of wireless
communications. In the wired world, each network link is independent and an
abstraction. In the wireless world, they are an interconnected ecosystem, where
each action by one has some impact on the others. The issues of spectrum deconfliction, nonlinear receiver responses, sensing, and the other topics covered in this
chapter have no analogs in the wired world and thus will go unaddressed unless
the wireless community becomes more involved in extending its practice into the
more general networking problem space.
References
[1] G.J. Foschini, “Layered Space Time Architecture for Wireless Communications in
a Fading Environment Using Multiple Antennas,” Bell Labs Technical Journal,
Vol. 1, No. 2, Autumn 1996, pp. 41–59.
182
Spectrum Awareness
[2] G.J. Foschini and M.J. Gans, “On limits of Wireless Communications in a Fading
Environment When Using Multiple Antennas,” Wireless Personal Communications,
Vol. 6, No.3, March 1998, p. 311.
[3] http://www.nari.ee.ethz.ch/commth/pubs/p/proc03
[4] http://www.pimrc2005.de/Conferences_en/PIMRC+2005/Tutorials/T4.htm
[5] http://userver.ftw.at/zemen/MIMO.html
[6] A. Gershman, Space-Time Processing for MIMO Communications, Wiley,
May 2005.
[7] S. Sharma and A. Nix, “Dynamic W-CDMA Network Planning Using Mobile
Location,” Vehicular Technology Conference, 2002 Proceedings, VTC 2002, Fall
2002, September 24, 2002, pp. 182–1186.
[8] C. Tschudin, “Fraglets—A Metabolistic Execution Model for Communication
Protocols,” in Proceedings of the 2nd Annual Symposium on Autonomous Inelligent
Networks and Systems, Menlo Park, 2003.
[9] C. Tschudin and L. Tamamoto, “Self-Healing Protocol Implementations,” Dagstuhl
Seminar Proceedings, 24 March 2005.
[10] http://www.cs.ucsb.edu/htzheng/cognitive/
[11] Q. Wang and H. Zheng, “Route and Spectrum Selection in Dynamic Spectrum
Networks,” Dyspan Conference, Baltimore, November 2005.
[12] P. Kyasanur and N. Vaidya, “Protocol Design Challenges for Multi-hop Dynamic
Spectrum Access Networks,” Dyspan Conference, Baltimore, November 2005.
[13] V. Naware and L. Tong, “Cross Layer Design for Multiaccess Communications
over Rayleigh Fading Channels,” IEEE Transactions on Information Theory and
Networking, submitted March 2005.
[14] V. Naware and L. Tong, “Smart Antennas, Dumb Scheduling for MAC,” in
Proceedings of the 2003 Conference on Information Sciences and Systems,
Baltimore, March 2003.
[15] Federal Communications Commission, Spectrum Policy Task Force Report,
15 November 2002.
Appendix: Propagation Energy Loss
Radio waves propagating in free space lose energy over the distance that the signal travels. This loss of energy is called propagation loss. It is usually measured in
decibels (dB) of loss on a logarithmic scale, in which 3 dB means half as much
power is received as was transmitted, 6 dB means one-fourth as much power,
10 dB means one-tenth the power, 20 dB means one-hundredth the power, 30 dB
means one-thousandth as much power, and so forth. Propagation loss may commonly range up to 100 dB (one ten-billionth the power received as transmitted) for
practical communication ranges.
183
Chapter 5
dB 10 Log10(power)
Free space power loss can be easily understood with the example of a light bulb.
Consider the light falling on the inner surface of a sphere surrounding the light
bulb at a distance 1 m from the light bulb. With that radius, the surface area of the
sphere is A 4R2 12.566 m2. So the percentage of the light falling on 1 m2 is
(1/12.566) 7.96 percent of the energy transmitted. At twice the distance (or
radius), the surface area of the sphere is four times larger, so the amount of energy
falling on 1 m2 is one-fourth as large; in the case of radio wave propagation, that
would be called a 6 dB loss. The antenna of a radio receives the transmitted radio
energy much as the light on an area on the inner surface of a sphere. Therefore, in
free space the power received falls another 6 dB at each doubling of transmitted
distance. Because propagation loss in this condition is proportional to the propagation range squared, this is usually referred to as R2 conditions.
Radio waves propagating across the surface of Earth, however, lose energy
more rapidly than in free space. This occurs in part because the signals bounce off
Earth’s surface and back up to the receive antenna, but have traveled a different
distance. At certain distances where the reflected signal has traveled exactly a half
wavelength farther than the signal arriving without a reflection, the signals add
together but at opposite phase and cancel each other, resulting in a very high propagation loss. In addition, trees and buildings absorb and reflect the signal propagating along the surface of Earth as well. As a result, propagation losses along the
surface of Earth increase not as the square of distance, but as the third or even
fourth power of distance, resulting in a 9 dB or even a 12 dB increase of propagation loss per doubling of range. This is commonly called R3 and R4 conditions,
respectively. To account for the average propagation loss in suburban conditions,
the industry often chooses to average the loss per doubling of range, and refers to
this as R3.8 propagation conditions.
184
CHAPTER 6
Cognitive Policy Engines
Robert J. Wellington
Department of Physics, University of Minnesota
Bloomington, MN, USA
6.1 The Promise of Policy Management for Radios
In familiar usage, policies are procedural statements expressing administrative
conventions that are adopted by various organizational entities. The concept of
automatic policy management of resources has its commercial roots in the administration of information systems and networks. Policy management refers to a particular approach for automating network management activities by specifying
organizational objectives that can be interpreted and enforced by the network
itself. The automatic application of management policies provides flexibility to
change the configuration of network devices at run-time to satisfy administrative
goals and constraints regarding security, resource allocation, application priorities,
or quality of service (QoS). A “policy engine” is a program or process that is able
to ingest machine-readable policies and apply them to a particular problem
domain to constrain the behavior of network resources.
This chapter concerns the application of policy management to cognitive radio
technology in general, and to spectrum management for frequency-agile radios in
particular. It focuses on what lessons can be learned from prior applications of
policy management to network resource problems. It reviews and leverages previous standards, research, and commercial implementations for policy engines and
applies them to the architecture and design of policy engines for cognitive radios.
6.2 Background and Definitions
The policy engine is the main inference component that triggers responses to
events that require changing the resource configuration. Often the output of the
185
Chapter 6
policy engine amounts to configuration commands or authorizations that are
tailored to specific kinds of network devices. In this sense, the policy engine
bridges the gap between domain-specific objectives and device-specific
capabilities. Despite the intrinsic need for interfacing with particular vendor
devices, a popular research trend has been to postulate the policy engine as a
general-purpose tool capable of deductive reasoning based on rules. Seen in
this manner, policy engines are descendents of the rule-based programming
frameworks that were popular in the 1980s. Bemmel et al. [1] describe an
expert system as simulating human reasoning using heuristic deduction rules,
where knowledge is stored as facts and new facts are derived by using a set of
deduction rules.
Much of the research in the area of policy-based networking has focused on
the specification of formal languages for expressing complex policies for various
domains and network management problems. Policies are expressed as sets of
rules about how to change the behavior of the network. Chadha et al. [2] define a
policy to be “a persistent specification of an objective to be achieved or a set of
actions to be performed in the future or as an on-going regular activity.” Carey
et al. [3] explain that “policies are expressed in terms of an event that triggers the
evaluation of a policy rule, a set of conditions that must be met prior to changing
the behavior, and a set of actions that are performed to change the behavior.”
Two trends are evident in the literature: (1) the deconstruction of policies into sets
of conditional rules of varying degrees of complexity, and (2) the use of objectoriented representations to support machine readability.
Figure 6.1 calls out commonly recognized functions and relations in a conceptual architecture for a network policy management system. The interpretation and
application of these functions for cognitive radio networks requires revisions to
this conceptual model, which are explored in this chapter. Evidently, the network
resource is the cognitive radio, and the policy decision point (PDP) and policy
enforcement point (PEP) represent new functions that enable policy management
of cognitive radio networks. The purpose of the PDP and PEP functions are to
interpret policies to control the behavior of the network devices to satisfy both the
users and administrators of network resources.
To implement the policy management architecture shown in Figure 6.1, the
system must monitor real-time network events and trigger the policy engine to
decide how current device states (policy conditions) should be mapped into
desired policy actions that can be quickly enforced by controlling specific device
operations. Carey et al. [3] note that policies allow less centralized and more flexible management architectures by enabling administrative decisions to be made
186
Cognitive Policy Engines
Policy
Administration
Tool Kit
Disseminate Policy
and Coordinate
Define Policies
and Select
System
Administration
Manage
Resources
Constrain
and Permit
Network
Administration
Tool Kit
Configure
and Status
Policy Enforcement
Point (PEP)
Monitor
and Control
Request and
Deliver Service
User
Legend:
Policy Decision
Point (PDP)
Network
Resource
Real-Time
Non-RT
Figure 6.1: Policy management system concept.
closer to where the event and conditions are actually detected. We must examine
how this approach applies to the particular problem of spectrum management for
cognitive radios.
6.3 Spectrum Policy
The usable frequency range of radio spectrum is divided into frequency bands
called allocations for particular types of use. The US frequency allocations are
available online [4]. Within a particular allocation, an allotment is a frequency
channel designated for a particular user group or service in some country. An
assignment is a license that grants authority to a specific party to operate a transmitter on a specific channel under specific conditions. The allotments and assignments are associated with a particular geographic area.
Historically, the allotments for broadcast services have been deconflicted
to ensure that the signal strengths in one area will not create interference for
signals in another area. However, it is not easy to define interference unless certain assumptions are made about the capability of radio receivers to reject interference in adjacent bands. Other technical considerations will include uncertain
187
Chapter 6
propagation characteristics, locations of nearby receivers or transmitters, and limitations of waveforms.
6.3.1 Management of Spectrum Policy
The International Telecommunication Union (ITU) Radiocommunication sector
[5] plays a global role in the management of the radio frequency (RF) spectrum
and satellite orbits. Around the world, RF spectrum is considered a finite natural
resource that is increasingly in demand from a large number of services, such as
fixed, mobile, broadcasting, amateur, space research, meteorology, global positioning systems (GPS), environmental monitoring, and last but not least, those
communication services that ensure safety of life at sea and in the skies. World
Radiocommunication Conferences (WRCs) are held every 2-3 years to review,
and, if necessary, revise the Radio Regulations, the international treaty governing
the use of the RF spectrum.
Within the United States, the Federal Communications Commission (FCC)
decides spectrum policy for commercial radio communications, and the National
Telecommunications and Information Administration (NTIA) plays a complementary role for the federal sector. Due to the perceived spectrum shortage resulting
from prior policies and the burgeoning demand, there has been great interest in
rethinking the leasing and allocation of spectrum. The FCC’s Spectrum Policy
Task Force (SPTF) [6] reported in 2002 that “spectrum policy is not keeping pace
with the relentless spectrum demands of the market.” Two recurring recommendations have been to “migrate from the current command and control model to [a]
market-oriented exclusive rights model and unlicensed device/commons model”
and to “implement a new paradigm for interference protection.” The FCC intends
to facilitate cognitive radio technologies to this end. In particular, the discussion
has often focused on “spectrum enhancing technologies,” including softwaredefined radios, leasing certain spectrum bands or “white space,” and sensory or
adaptive devices that could find unused spectrum. For example, to identify unused
RF channels, a cognitive radio might be able to measure an “interference temperature” for the ambient spectral environment.
Deciding which policies apply to a particular cognitive radio will require
understanding the role the radio is playing for the user of some service at a particular location. The policy language must be rich enough to express the semantics
of the possible spectrum management policies, and the cognitive policy engine
must be able to automatically interpret and enforce the applicable policies. The
FCC and NTIA are contemplating pilot projects to explore options for encoding
188
Cognitive Policy Engines
selected spectrum policies in a machine-readable language. Section 6.4 looks
deeper into these languages.
6.3.2 System Requirements for Spectrum Policy Management
A comprehensive policy management system that could autonomously control
interference created by frequency-agile radios would have to include an extensive
syntactic capability to specify policies and policy engines that can interpret policy
semantics expressing technical concepts such as authorization, frequency bands,
channels, propagation conditions, signals and noise, waveforms, geopolitical
boundaries, geographic locations, dates, times, types of services, and possible
roles for the various radios in the environment.
The policy engine must be able to automatically modify the run-time configuration of the cognitive radio to sense the spectrum environment; to detect other
local radio networks; and to control its own transmissions by adjusting frequency,
power, modulation, signal timing, data rate, coding rates, and antenna.1 In summary, the spectrum policy management system for cognitive radios shall:
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
Be distributed across multiple radio platforms with autonomy at the radio level.
Permit network administrators to determine applicable spectrum policies.
Represent spectrum policy rules in a machine-readable format.
Resolve conflicts and inconsistencies in sets of policy rules.
Designate specific roles and services for each radio.
Require authorization for policy changes or updates for radio.
Monitor spectrum utilization and detect RF interference.
Control run-time configuration of cognitive radios to satisfy spectrum policies.
Support heterogeneity and diversity of legacy vendor radios.
Support continual growth and increasing complexity of cognitive radios.
6.4 Antecedents for Cognitive Policy Management
The last decade has seen significant research and applications for policy management technology in the area of network management. Tacit assumptions about the
1
Multiple policy engines—one for the physical layer, one for the network, and one for the user—
are possible; this chapter mainly covers the physical layer policy engine. In addition, there may be
engines for equipment-specific implementation, behavior-specific policies, or other policies.
189
Chapter 6
design of the Internet and related information technology (IT) infrastructures have
greatly influenced the development of policy management approaches. It is not
surprising that much of the research has focused on the limitations of descriptive
ontologies for web-enabled applications and packet network resources. However,
Kagal et al. [7] rightly caution that reasoning about policies generally requires
application-specific information, forcing researchers to create policy languages
that are bound to the domains for which they were developed.
Chadha et al. [2] have surveyed policy-based network and distributed systems
management approaches that have been the subject of extensive research over the
last decade (see also [8]). The Internet Engineering Task Force (IETF) [9] has
sponsored standardization efforts for object-oriented models for representing policies
as well as a framework and protocols for managing Internet Protocol (IP) networks.
This section examines how well the existing state of the art can be adapted
to support the performance characteristics of radio networks and the specialized
requirements for spectrum resource management. This section draws from government projects, commercial applications, academic research, and standardization
efforts that provide a context for designing a cognitive policy engine specialized
for spectrum management.
6.4.1 Defense Advanced Research Projects Agency Policy Management Projects
The Defense Advanced Research Projects Agency (DARPA) has funded research
and development (R&D) efforts that push forward the technology frontiers in the
area of network policy management in general and even the particular application
to spectrum management. This section reviews some of that work that is in the
public domain.
Funding for contractors involved in the DARPA NeXt Generation (XG) radio
communications program [10] has been a very significant, if not the principal, driving force behind the development of policy management techniques for radios. The
FCC now has complementary projects to define policies for frequency-agile radios.
The XG program followed on the heels of the DARPA Policy-Based Survivable
(PolySurv) communications program, which demonstrated that increased military
survivability and real communication performance gains could be brought about by
downloading dynamic mission policies to automatically manage radio networks.
XG is now focused on system concepts and enabling technology to dynamically
redistribute allocated spectrum in operating radio networks in order to address rapidly growing requirements for communications bandwidth. The program goals are
to enable radios to automatically select spectrum and operating modes in a manner
190
Cognitive Policy Engines
that increases the survivability of communication networks and minimizes disruption to existing users.
The DARPA Dynamic Coalitions program [11] has strongly influenced the
technical approach for XG by offering policy representations and policy engines
that support other network management techniques. For example, Phillips et al.
[12] describe constraint-based models and the application of role-based access
control (RBAC) for implementing security policies in the context of Dynamic
Coalitions. Uszok et al. [13] describe a Semantic Web (SeW) language [14] called
KAoS that has proliferated with support from DARPA. In fact, KAoS was based on
the DARPA Agent Markup Language (DAML) [15], and KAoS has capabilities
that support both the expression and enforcement of policies in a software agent
context. The policy language has always been based on the eXtensible Markup
Language (XML) to support common Web services, but due to shortcomings of the
DAML description logic, KAoS now relies on the Web Ontology Language (OWL)
[16] to represent knowledge about domains and rule-based policies. In the KAoS
environment [17], domain managers act as PDPs and are responsible for administering policy for entire domains.
DARPA has also been active in funding more traditional network policy
management approaches for the Next Generation Internet (NGI), and Stone et al.
[18] provide background information that is relevant for policy management of
cognitive radio networks.
6.4.2 Academic Research in Policy Management
This section looks at what a spectrum management implementation might leverage from the research community concerning existing policy languages and
frameworks for network policy management. Carey et al. [3] include an overview
of the state of the art in policy languages that addresses access control and
resource management. Rules governing spectrum access can be enforced by a
radio that subjects itself to access control policies. In computer network management systems, only users associated with accounts included in an access control
list (ACL) can access the resource. The next enhancement is association of users
with groups and roles, leading to RBAC systems. Spectrum resources are already
assigned to particular radio services, so the cognitive policy engine can process
attributes associated with roles for the radio to provide a context for evaluating
spectrum access control rules. Ideally, we would want users to authenticate themselves to the radio and associate different types of authentication with different
roles for the user, the radio, the network, and network resources.
191
Chapter 6
Strauss [19] describes the requirements and architecture of a policy management system based on the IETF script management information base (MIB) infrastructure. An MIB is a device-specific database for remotely managing a network
resource using the Simple Network Management Protocol (SNMP) based on IP
communications. Most IP-capable devices support an MIB that provides a standard
interface to monitor and configure the device. For the cognitive radio, we could
envision a “spectrum MIB” with a standard interface supported by an underlying
device-specific mechanism to actually configure the network elements. Strauss [19]
used this approach to implement network QoS control with the Jasmin Script MIB
agent [20]. In this case, the policy engine is just the Java run-time engine
executing policy scripts supported by policy-class libraries for different device
capabilities.
Ponder is a different policy management framework that includes a relatively
mature policy language with a suite of tools and source code that has been freely
available to download from the Internet [21]. Ponder tools support administration
of domain hierarchies, positive and negative authorization policies, delegation
policies, and event-triggered condition-action rules. Dulay et al. [22] describe how
to use the Ponder framework to encode, disseminate, and process security and
management policies for distributed applications. In fact, Ponder would be an
initial starting point for developing an administrative toolkit that could be used
to specify, compile, maintain, and disseminate spectrum policies for cognitive
radios. It integrates with a domain server and supports role abstractions that could
be used to manage spectrum policies for multiple communications services that
might eventually be supported by cognitive radios.
Ponder has been well tested in various applications [23], and “back-ends” (i.e.,
application-specific PDP and PEP functions) have been implemented to generate
firewall rules, Windows access control templates, Java security policies, and obligation policies for a policy agent. The Ponder language for representing policies
has been described as a declarative, object-oriented language that can express both
“obligation” and “authorization” policies [24]. To be specific, Damianou et al. [23]
explain that “policies define choices in behavior in terms of the conditions under
which predefined operations or actions can be invoked rather than changing the
functionality of the actual operation themselves.” Obligation policies are defined as
“event-triggered condition-action rules that can be used to define adaptable management actions,” and authorization policies are “used to define what services or
resources a subject (user or role) can access.”
The policy research community in general is particularly concerned about the
difficult task of analyzing the meaning of groups of policies to determine the
192
Cognitive Policy Engines
implications for particular agents and resolving possible conflicts between policies. Even if spectrum management objectives are clearly stated in a policy, the
implications for device configurations or required actions are not always obvious
in practice. For example, consider a long-standing policy that authorizes a particular frequency band for some type of messaging service, and specifies servicespecific protocols for users to share airtime. Suppose a newer policy for cognitive
radios specifically authorizes a class of cognitive radio users to share an overlapping band of spectrum subject to different limitations on availability of channels
for legacy users. Is it clear what the airtime restrictions are for a particular cognitive device that performs a similar type of messaging using the legacy channels
and protocols? Which rule takes priority, or must both usage restrictions be
observed by the cognitive radio? Do permissions take precedence over prohibitions? Policy refinement is the process of deriving lower-level, more specific policies that the device can enforce in order to completely meet the requirements of a
group of management policies.
Damianou et al. [24] describe other problems with policy refinement, and
Ponder provides tools for policy analysis and refinement to assist administrators
in detecting and resolving policy conflicts. In particular, Ponder supports the
introduction of priorities and preferences. A simple method to resolve policy
conflicts for a device is to assign explicit priorities to every policy so it is clear
which policies overrule others. Locally, rules can be prioritized for the device in
order to reflect local management priorities, such as ensuring that efficiency is
more important than reliability, or vice versa. For a specific device, sets of rules
are also ordered by update times, particularly if partial updates of the rule base are
accepted practice. DAML relies on update times as well as numeric priorities to
determine priority [24].
Stone et al. [18] suggest the idea of differentiating policies “by their granularity, such as the application level, user level, class level, or service level,” and
letting spectrum managers designate certain mission-applications for priority.
Hierarchical policy management and domain groupings provide another degree
of flexibility, permitting a PDP to branch beyond the linear ordering of priorities.
The device may give priorities to policies originating within a more local domain,
given an inheritance hierarchy. Similar to the manner in which federal policies
overrule state policies, Uszok et al. [17] anticipate a “policy harmonization”
process that invalidates portions of the lower-priority policy to resolve the conflict. Decisions about how a PDP should handle inheritance must be made at the
time that the policy hierarchy is established. In addition, these decisions should
belong to the human realm of policy administration. Experience tells us that it
193
Chapter 6
takes a human judge to decide how to invalidate portions of state policies to eliminate conflicts with federal rules.
Another way that a PDP can handle nonlinear priorities and resolve conflicts
and ambiguities between overlapping rules is to permit the specification of “metapolicies” constraining the interpretation of groups of policies. For example, Kagal
et al. [7] describe another policy language, Rei, that supports meta-policies for
conflict resolution. Rei was designed for general application and permits domainspecific information to be added without modification. Tonti et al. [25] provide
a comparison of the capabilities and shortcomings of KAoS, Ponder, and Rei,
rendering a valuable perspective of the various approaches.
Clearly, the techniques for resolving policy conflicts could be a fruitful area of
research for a long time to come. As far as spectrum management is concerned, it
appears that the research community has already done enough work to get started
expressing FCC policies in a machine-readable format. A number of abstract policy languages already exist, and the task ahead is to introduce terminology that is
directly applicable to spectrum management and the cognitive radio domain. The
challenge will be to come up with a reasonably useful and self-consistent rule
base for cognitive radios that does not present an opportunity for litigation about
the meaning, implications, or applications of the rules.
6.4.3 Commercial Applications of Policy Management
This section examines what can be learned from commercial products that could
be useful for designing cognitive policy engines. No one will be surprised to discover that such major network vendors such as Cisco, Nortel, and Lucent Technologies have already developed policy management products to support
administration of local and wide area networks (LAN and WAN).
Damianou et al. [23] indicate that Lucent has used a Policy Definition Language
(PDL), similar to Ponder, to program Lucent switching products. PDL “uses the
event-condition-action (ECA) rule paradigm of active databases to define a policy as
a function that maps a series of events into a set of actions.” This approach is interesting because in addition to policy rules, there are policy-defined event propositions that allow groups of simple events to trigger more complex events.
Nortel [26] advertises its Optivity Policy Services with system components
consisting of a policy server, management console, directory server, and policyenabled network components. This distributed IP network architecture interoperates using Lightweight Directory Access Protocol (LDAP), Common Open Policy
Service (COPS), and Command Line Interface (CLI). The administrator is
194
Cognitive Policy Engines
required to select from a number of predefined policies (i.e., “if condition, then
action” rules) that relate to the roles of various network devices identified in the
directory server. The policy server then issues COPS or CLI commands to configure devices such as routers, IP telephone gateways, and firewalls.
Damianou et al. [23] observe that a common component of commercial tools
is a graphical user interface (GUI), which typically allows the administrator to
visually select a network device or other managed element from a hierarchically
arranged tree-view of policy targets, and specify the policies in the form of “if
conditions, then action rules for the selected targets.”
What comes across in all the commercial examples is that the network devices
must be designed to support policy rules that can configure their behavior by some
mechanism. The sophistication of the devices determines the nature of the interface
with a policy server that either directly issues device configuration commands,
supports a COPS dialog to make policy decisions for the device, or disseminates
defined policies that are recognized by the device. Ultimately, the nuances of the
policy language seem to be relatively unimportant compared to the sophistication
of the network devices.
6.4.4 Standardization Efforts for Policy Management
The proliferation of vendor architectures for policy management of telecommunications networks has motivated the IETF to address standards for interoperability.
Using the Policy Core Information Model (PCIM), Moore et al. [27] begin the
process of standardizing policy management terminology and representations of
network management policies to provide an accepted framework for vendorspecific implementations.
To what extent can these standardization efforts be applied to policy engines
for spectrum resource management for cognitive radio operations? Initial cognitive radio implementations will necessarily focus on radio-specific features, and as
the technology proves successful, the focus will extend to larger radio networks
rather than individual radios. The IETF policy domain is already oriented toward
management of large-scale telecommunications networks, with the goal to assist
human network administrators to define machine-readable policies and architectures for the automatic control of network resources. The cognitive radio is analogous to a particular network device.
Snir et al. [28] envision a physical network architecture in which the PDP
translates abstract policy constructs into configuration commands for multiple
devices (e.g., a router, switch, or hub) where policy decisions are actually
195
Chapter 6
enforced (i.e., PEP). Although there may be compelling arguments for the architectural assumption that one PDP services multiple PEPs in the case of high-speed,
high-reliability network architectures, this is not so clear for the cognitive radio
application. Stone et al. [18] point out that an underlying assumption of PCIM is
that policies are stored in a centralized repository, and the PDP is the entity in the
network where policy decisions are made using information retrieved from policy
repositories. When the PEP requires a policy decision about a new flow of traffic
or authentication, for example, “the PEP will send a request to a PDP that may
reside on a remote server.”
Moore et al. [27] indicate the PCIM standard fits into an overall framework for
representing, deploying, and managing policies that is being developed by the
IETF Policy Framework Working Group. In Figure 6.1, the link between the PDP
and PEP has two characteristics: (1) it needs to operate in near real-time for
timely enforcement decisions and (2) it is conceived to be a query–response dialog. For the cognitive radio application, due to concerns about link reliability and
bandwidth, the first assumption is tenable only if the PDP and PEP functions are
colocated on the radio platform. Furthermore, the query–response design is also
very natural, given the prevalence of client–server and three-tier database transaction architectures in the Internet. In fact, two important Internet applications for
policy management involve access control (i.e., security) and admission control
(i.e., QoS), and both involve permission. However, there is no reason to assume
this Internet design is optimal for a cognitive radio application in which the PDP
and PEP functions are colocated.
In PCIM, the policy-controlled network is modeled as a state machine, in
which policy rules control which device states are allowed at any given time. Each
policy rule consists of a set of conditions and a set of actions. Policy conditions are
constructs that can select states according to complex Boolean expressions. Policy
actions are device behaviors, such as selecting or prohibiting certain frequency
bands, bandwidths, protocols, coding, or data rates. When events lead to certain
conditions, then certain actions can or must be performed by the device (depending
on whether the actions are obligated or simply permitted for the device).
Bemmel et al. [1] describe these policy rules as examples of ECA rules, by
which changes within a system or the environment trigger adaptation of the system’s behavior. Such systems are called event-driven, and many programming languages and environments support compatible software development techniques.
For example, in the Microsoft Windows® architecture, events are basically messages that are routed between processes called event handlers. Unfortunately, if
the policy rules were coded in this familiar manner, the component’s behavior
196
Cognitive Policy Engines
would basically be hard-coded when the program is compiled. We need a more
flexible way, however, to bind actions to the event handlers at run-time for example,
Java Remote Method Invocation (Java RMI) or Common Object Request Broker
Architecture (CORBA).
The PCIM defines class representations and abstract attributes for policies, but
not the algorithms or design of the policy engine. Figure 6.2 depicts the ontology
for the object-oriented design of policy classes in the PCIM. Policy rules are
aggregated into policy group classes. These groups may be nested to represent a
hierarchy of policies. Although retaining all the same attributes of the policy
classes is not particularly important, a similar class diagram will be suitable for
representing policies for the cognitive radios.
Policy
Common Name : string
Policy Keywords : string
Policy Rule
Enabled : unit 16
Condition List Type : unit 16
Rule Usage : string
Priority : unit 16
Madatory : boolean
Sequenced Actions : unit 16
Policy Roles : string
Policy Group
Policy Time
Period Condition
Time period : string
Month of Year Mask
: octet string
Day of Month Mask
: octet string
Day of Week Mask
: octet string
Time of Day Mask : string
Local or utc time : unit 16
Policy Condition
Vendor Policy
Condition
Constraint [] : octet string
Constraint Encoding : string
Policy Action
Vendor Policy
Action
Action Data [] :
octet string
Action Encoding : string
Figure 6.2: Unified Modeling Language (UML) diagram for PCIM policy classes.
One important attribute of the policy rule is the ability to associate a “role” for
the radio. The policies (e.g., assigned frequency bands) for land mobile radios
(LMRs) are different from those for air traffic control (ATC). Moore et al. [27]
197
Chapter 6
stress that rather than configuring—and then later having to update the configuration of—hundreds or thousands (or more) of resources in a network, a policy
administrator assigns each resource to one or more roles, and then specifies the
policies for each of these roles. The policy framework is then responsible for configuring each of the resources associated with a role in such a way that it behaves
according to the policies specified for that role. When network behavior must be
changed, the policy administrator can perform a single update to the policy for a
role, and the policy framework will ensure that the necessary configuration
updates are performed on all the resources playing that role.
6.5 Policy Engine Architectures for Radio
6.5.1 Concept for Policy Engine Operations
This section begins to synthesize a concept for how the cognitive policy engine
will operate. Section 6.4 shows how network policy management systems operate
in a distributed fashion with administrators using policy languages to express
network objectives, policy servers that act as PDPs deciding how to configure
devices to achieve these goals, and policy-enabled devices that are configured to
enforce the policy (i.e., PEP). The spectrum management objectives for the cognitive radio are slightly different than network access and resource control, for
which the primary goals are security and network application performance. The
primary interest here is ensuring that the frequency-agile radio is utilizing available spectrum resources in accordance with regulatory licenses and prohibitions
while acting as a good neighbor to avoid interference with other users. So there is
a need to distinguish between a PDP that can optimize global network performance, and a PEP that configures the cognitive radio to obey locally enforceable
policy constraints. In other words, the cognitive policy engine is responsible for
device configuration.
The relationship between the PDP and PEP functions is also different for the
cognitive radio. Carey et al. [3] describe a typical network policy management
dialog in which events detected by the device cause the PEP to formulate a
request for a policy decision and send it to the PDP (usually a policy server), such
as when Resource ReSerVation Protocol (RSVP) is employed by a user application to advertise its QoS requirements. COPS may be used as a policy transaction
protocol between the PDP and PEP for transporting the policy requests and decisions. The PDP returns the policy decision and the PEP then enforces the policy
decision and responds to the RSVP from the user-application accordingly.
198
Cognitive Policy Engines
In our concept, the cognitive policy engine will perform both PDP and PEP
functions for the radio using local platform resources. The operation will generally
accord with Figure 6.1, and involve the steps shown in Table 6.1. There is no necessity for radio device functions to request a policy decision from the policy engine
because the policy engine is already responsible for enforcing the policy on the platform. As shown in Figure 6.1, the policy functions (PDP and PEP) act to monitor
and control the radio platform to satisfy any policy constraints or obligations.
Table 6.1: Concept of operation for the cognitive policy engine.
Step 1
The policy engine receives a download of policies to be observed by the radio.
Step 2
The policy engine determines what information to monitor on the radio
platform (e.g., location, time)
Step 3
The cognitive radio provides the requested situation data (e.g.,
location, time) to the policy engine as required
Step 4
The arrival of new information is an event triggering some processing of
policy rules
Step 5
The policy engine formulates applicable constraints on spectrum usage for
the radio
Step 6
The policy engine configures the radio to observe spectrum constraints and
perform obligatory activities (e.g., communication protocols)
Step 7
The cognitive radio operates within defined constraints and performs
obligatory activities
Step 8
The policy engine performs any obligatory hierarchical reporting activities
for network management
If we assume the cognitive radio is administered by a central policy authority
in a domain that recognizes the capabilities of the particular cognitive radio and
disseminates all enforceable policies directly to the cognitive radio, then the
complexity of the policy engine can be reduced. For example, the policy engine
does not have to determine whether the policies can be enforced on the platform.
Any ECA rule sent to the platform could (but not should!) even be compiled and
executed by a Java engine on the radio.
In the case of radio networks, such as an Internet community, no centralized
authority will own all of the resources. It will not even be possible to guarantee
efficient connectivity among the worldwide users of cognitive radio technology.
Feeney et al. [29] argue against a centralized policy administrative authority in
199
Chapter 6
cases of great organizational diversity, and propose a concept of operations with
hierarchies of policy authorities. This approach reflects a real-world community
of users, but it leads to the possibility of policy conflicts that must be resolved
among organizations. These conflicts must be resolved at the network level, and
the cognitive radio should not be forced to handle this difficult problem.
6.5.2 Technical Approaches for Policy Management
This section proposes a technical approach for designing the cognitive policy
engine to satisfy the concept of operations presented so far in this chapter. We
look first at what must be done to design the policy language and then examine
alternatives for the technology behind the policy engine.
Multiple approaches for policy specification have been proposed that range
from formal policy languages that can be processed and interpreted easily and
directly by a computer, to rule-based policy notation using an if-then-else format,
to the representation of policies as entries in a table consisting of multiple attributes [30]. This chapter has already looked at alternatives, including the compiled
policies of Ponder, Java scripts supported by interpreted classes, KAoS with its
OWL semantics, RBAC, Rei, and PDL.
Whatever language is selected, the syntax will have to be extended to support
the technical jargon of spectrum policy with such attributes as frequency bands,
channels, propagation conditions, signals and noise, waveforms, geopolitical
boundaries, geographic locations, dates, times, and types of services. The language
should support authorization and obligation policies, and roles for the various
radios in the environment. In addition, the language should make a clear distinction between management policies and the resources and activities being managed
[31]. The plan from the beginning should consider eventual “use of domains as a
means for grouping resources with dependencies reflecting both hierarchical interactions (e.g., control of resources, authority delegation to subordinate managers)
as well as peer-to-peer interactions (e.g., negotiations between peer managers to
prevent/resolve management conflicts)” [31].
Looking forward, the ontology should support standardization efforts for multiple vendors of cognitive radios, even though it will likely be redesigned in the
future. For this reason, we should eschew compiled and interpreted languages,
and start with the flexibility provided by SeW languages such as OWL. Then the
ontology will be built up as capabilities and understanding increase over time.
The choice to start with a SeW language does not exclude the use of compilers
and interpreters; it just focuses this technology on the implementation of actions
200
Cognitive Policy Engines
and behaviors with processes named in the policies. For example, useful radio
protocols should be named in the ontology, and policies relating to standard protocols should be defined. As more device-specific actions are represented with new
terminology, the generality of the policy architecture will decrease because the
language must support increasingly complex and specialized actions for proprietary
device behaviors [32]. Another shortcoming of this approach is that the policy engine
and ontology have difficulty reasoning about complicated, composite behaviors.
Selecting the right balance between generality and specificity in referencing cognitive radio behavior is an area for further research that need not obstruct initial
efforts to create a spectrum policy language for the policy engine.
For policy engine development, there are also several approaches to consider.
Note that the engine itself need not internally use the policy language as its input,
because the policies can be interpreted when they are downloaded to the cognitive
radio. Named behaviors can be bound in the radio to specific procedures and algorithms. For example, a compiler or just-in-time interpreter can create “tokens” from
the input and link together whatever processes or logical structures are used internally to the cognitive radio.
Turning to the technical approach for the cognitive policy engine, it is important to recognize that the policy engine will be performing both PDP and PEP
functions on the same platform. The concept of operation proposed here is that the
policy engine will be responsible for both interpreting and enforcing spectrum
policies, as well as monitoring platform events to trigger changes in the configuration of the radio functions. Again, according to Figure 6.1, the functional interface
between the policy engine and the radio platform is defined as a relationship of
monitoring and controlling the radio. This is the situation shown in Figure 6.3(a).
Figure 6.3: Alternative policy
engine architectures [SR2].
Policy
Database
Current
Policy
Policy
Engine
Policy
Constraints
(a)
Policy
Database
Current
Policy
Current
State
Policy
Engine
Current
State
Proposed
Configuration
Regulatory
Validation
(b)
201
Status
Monitor
Radio
Control
Status
Monitor
Radio
Control
Chapter 6
An alternative approach characterizes the purpose of the interface as validating
that the operation of the radio complies with all relevant policies. This situation is
shown in Figure 6.3(b). Uszok et al. [17] argue that the interface between a PEP and
a native environment can be standardized to answer the question “is a given action
authorized or not?” This approach posits a query–response dialog between the policy engine and the native device. Thus the policy engine explicitly polices actions by
the radio, requiring the radio to create objects that describe proposed actions so that
the policy engine can pass judgment on whether the action is permissible.
Some involved parties (but not members of the FCC) have publicly voiced the
opinion that validation is the only technical approach that will support certification
of a frequency-agile radio by the FCC. Were this opinion correct, the policy engine
architecture would have to combine both functions in Figure 6.3. Function (a) performs vendor-specific configuration and control of a proprietary device and function (b) certifies policy compliance based on names for generic behaviors. In the
latter case, the policy engine can have only a partial understanding of the device’s
proprietary behavior. This leads to requirements for default authorization policies
in which actions that are not defined in the policy language should generally be
denied, or are assumed to have been deemed permissible when the radio platform
was certified by the FCC.
The technical approach presented here for the cognitive policy engine starts
with configuration rather than validation. The emphasis is on supporting obligation policies and ECA rules in addition to authorization policies typical of security
platforms. As Stone et al. [18] state, “a policy specifies what action(s) must be
taken when a set of associated conditions are met.” Bemmel et al. [1] consider this
policy-based approach with ECA rules to be “useful in cases when a controlled
part has many choices/options and a controller intervenes, enforcing a choice in
accordance with a particular policy goal.” This seems to be the likely situation,
as cognitive radio technology will increase in agility over time.
An important part of this approach at both the device level and the network level
is status monitoring. According to Lee et al. [33], “Feedback is a critical part of the
process where monitoring activities by the management system lead to changes in
the low-level policies.” This permits the policy management system to adjust the
radio according to changes in demand for bandwidth or interference conditions.
6.5.3 Enabling Technologies
This section completes the examination of what particular technologies will be
required to implement this technical approach. Carey et al. [3] summarize two
202
Cognitive Policy Engines
challenges for cognitive policy management: “Currently policy-based management
suffers from fragmentation of approach and there is no commonly accepted policy
language and no common approach to the engineering of policy based systems.”
Regarding the technology behind the policy engine, Khurana and Gligor [34]
approach the problem by modeling the network as a state machine and using policy to place limitations on the state transitions that are allowable at any given
time. Any model for the cognitive radio that is internal to the policy engine will
tend to become more complex and specialized over time. The same issue arose
with the increasing specialization of the policy language over time. More R&D
will determine whether this requires a custom design for the policy engine that
is matched to the radio’s capabilities.
The state transitions in the model can occur only when the constraints are
satisfied, and this may require logical deductions at some point. Damianou et al.
[23] note that logic-based policy languages have a well-understood formalism that
is amenable to analysis. For example, a policy engine supporting RBAC can be
implemented in the Prolog language because the policy constraints are equivalent
to restricted first-order predicate logic (RFOPL) statements [35]. Kagal et al. [7]
say that the Rei policy engine was developed in Java and used Prolog as a reasoning engine. Stone et al. [18] summarize the situation: “One clear method is to use
formal logic to represent network policies. Although this method would make
conflict detection much easier with the use of existing theorem-provers, most network policy implementers are not as comfortable with this representation.”
The policy engine behind KAoS is an online theorem-prover that permits logical reasoning about domains and policies [17]. The enforcement mechanisms are
Java classes that are specific to a resource platform, but capable of interpreting
and enforcing policies from the PDPs. The Java Theorem Prover (JTP) [36] supports queries consisting of properties defined in a given namespace and selection
of all possible values. The JTP will also provide a response regarding whether a
given assertion is true. For conditional policies, JTP can store a state consisting of
certain assertions that are true. These assertions can be removed when the condition is determined to be false by the radio or event monitor.
Another area for R&D is motivated by Carey et al. [3]: “Problems with policy
decomposition and hierarchical translation still present difficulties.” Damianou
et al. [23] characterize the situation: “Despite significant efforts in developing different policy specification techniques, the ability to refine such goals into concrete
policy specification would be useful. Policies can be used to support adaptability
at multiple levels in a network.” “Research is needed on defining interfaces for the
exchange of policies between these levels. It is not easy to map the semantics of
203
Chapter 6
the policies between the different levels.” The basic problem is that the higherlevel management applications “will not be aware of the network components and
cannot specify policies to be interpreted by them” [23].
At this time, the need for extensive logical reasoning capabilities in the policy
engine has been oversold, particularly at the level of the device. Whenever a
problem requires complex reasoning, it is generally solved with a special-purpose
algorithm or application. Two techniques are now available for use to help resolve
ambiguities in the rule base. First, policy rules can be associated with a priority
value to resolve conflicts between rules [24]. Second, “meta-policies” are policies
about policies, and may be used to detect semantic conflicts between policies [3].
Kagal et al. [7] describe this capability in Rei.
6.6 Integration of Policy Engines into Cognitive Radio
In the end, the integration of the policy engine into the cognitive radio is primarily
an issue of software system engineering. The process will involve deriving a complete set of software requirements from the system concepts; defining a software
architecture that is compatible with the cognitive radio platform software architecture; defining the software interfaces; and designing, coding, and testing the software components. Some assumptions must be made about the software architecture
and application-programming interface (API) provided by the services on the host
platform. This section, for example, makes the assumption that the first cognitive
radio implementation will likely evolve from the efforts of the Software Defined
Radio Forum (SDR Forum [37]). It begins with platform integration issues and
then moves on to issues related to integration into a policy-managed network.
6.6.1 Software Communications Architecture Integration
The Joint Tactical Radio System (JTRS) JPO defines the Software Communications Architecture (SCA) based on CORBA middleware [38]. The SCA is a layered architecture with well-defined APIs for integration with new applications
(called components). The applications are supported by a core framework that
provides interfaces for exchanging information, controlling software processes,
configuring the platform, and accessing files. The policy engine will be another
component in this architecture supported by the core framework with interfaces
to other components. One of the most important architectural considerations is the
definition of the interface between the policy engine and software environment of
the radio.
204
Cognitive Policy Engines
The policy engine can be divided into three main policy management functions: policy service, policy decisions, and policy enforcement:
●
●
●
The policy service maintains the policies that are downloaded to the radio frame
by a policy authority, and it requires an external network interface communication. The policy service also performs any necessary policy life cycle management operations, such as parsing the policy language, keeping track of whether
policies are enable or disabled, and removing policies if they become obsolete.
The policy decision function fulfills the requirement that the device have a local
PDP. To make these decisions, it needs information about the communication
conditions and spectrum events in the radio environment. Thus, it must interface to some kind of status monitor that can report conditions and real-time
events that require a policy decision.
The policy enforcement function acts as the PEP for the device, and it needs a
control interface to constrain the radio’s behavior and perform any obligated
actions demanded by policy.
Figure 6.4 shows how the policy engine functions are integrated into the SCA
environment by creating the necessary functional interfaces. The coupling
between the policy engine and the radio will be rather complex because there is
likely to be a big gap between abstract policy syntax and names, and because of
the detailed design of the radio software functions and interfaces. At least one of
the interfaces should be relatively simple to accomplish. The radio is already
capable of network communications, presumably through an API for data
Figure 6.4: Software
design for policy
management.
Cognitive Radio
Policy
Engine
Policy
Interface
SCA
Components
Policy
Service
Policy
Network
I/O
Component
Policy
Decisions
Policy
Enforcement
Status
Monitor
Waveform
Control
(New) Spectrum
Component
SDR Core Framework
205
Waveform
Components
Chapter 6
input/output (I/O), and a policy network interface can be implemented by software
that will interact with the SCA to securely download policies to the policy server.
The policy enforcement interface will be more complicated because the policy
actions will not generally correspond to particular controls for the attributes of the
waveforms that are being transmitted by the radio. The waveform control interface
must interact with the SCA waveform components to bridge this gap between
behavior specifications and control parameters. Depending on the type of policies
that will be enforced, it is likely that the waveform component may have to be
adapted to be “policy-enabled.” This adaptation would likely involve exposing
new API methods that were designed to support the actions required by the policies. For example, the waveform control interface will have to communicate any
limitations on transmission frequencies and power levels, limitations on protocols,
or times when transmission is permitted.
What is entirely new is the spectrum component, which is the focus of the status
monitor. The spectrum component will require signal-processing capabilities to
determine channel occupancy and any other information about radio conditions and
events that will be necessary to evaluate the ECA policy rules. The status monitor
interface must be able to “observe” the radio operation and its spectrum environment and maintain information about the state of the device. It will have to notify
the policy decision function about any relevant changes to the state that may trigger
policy rules. This functional interface must aggregate digital signal processing
results into higher-level event objects and provide event notifications to the policy
engine. Because the events and conditions of concern depend on the policy specification, the policy engine must inform the status monitor what radio and spectrum
attributes and conditions to observe, and what kind of notifications to return for consideration by the policy engine. Examples of the types of data that may flow across
this interface are current frequency band, current geographic location, current time
or data, and the communication roles of the radio in the network.
6.6.2 Policy Engine Design
Having discussed the external software interfaces for the policy engine, this section now looks deeper into the internal functional design of the policy engine by
looking at what others have done, proposing a top-level design, trying to define
software requirements, and building up a detailed design introducing new functionality as necessary.
Montanari et al. [39] describe a useful programming environment called
POEMA (Policy Enabled Mobile Applications), which gives one example of
206
Cognitive Policy Engines
designing policy-based “middleware.” Boutaba and Znaty [31] insist that a policy
engine design must have three logical components: execution engine, situation
matcher, and observer. The observer is responsible for defining the situation in
terms of “combinations of the values of observables (an attribute of some object
which has a value which can be measured) and times of observations.” The situation matcher makes policy decisions by “matching the observed state of the system with stored situation specifications” that correspond to the policy conditions
and policy events discussed previously.2 “The language used to express situations
must be able to represent observables and provide operators for obtaining the
value associated with an observation, and the time an observation was made” [31].
The execution engine operates on an “algorithmic block” (i.e., policy action) and
produces a “sequence of instructions” that the policy enforcement function discussed previously will use to control the radio.
Figure 6.5 offers a tentative functional design for the cognitive policy engine.
Due to the importance of the status monitor interface for the policy engine, it is
included in the figure because Boutaba and Znaty [31] consider it to be an integral
part of the policy engine. Note that two types of policy actions are produced by
the policy decisions function. Stone et al. [18] remark that “Policies can be triggered in two ways, either statically or dynamically.” “Static policies apply a fixed
set of actions in a predetermined way according to a set of predefined parameters
Figure 6.5: Functional design
for policy engine.
Policy Decisions
Active
ECA Rules
Policy Services
Parse
Syntax
Filter by Condition
Filtered
Rules
Trigger
Change
Constrain
Action
Policy
Actions
Policy Enforcement
Action
Semantics
2
Policy
Semantics
Behavior
Synthesis
Policy
Conditions
Policy
Events
Status Monitor
Waveform
State
Spectrum
State
The situation matcher may be implemented with a variety of technologies, ranging from complex
Boolean specification, to fuzzy logic, to neural nets, or even a Markov model.
207
Chapter 6
that determine how the policy is used.” In other words, these are ECA rules with
no policy events to trigger a change in policy. “Dynamic policies are only
enforced when needed, and are based on changing conditions … actions can be
triggered when an event causes a policy condition to be met” [18].
Some lessons about software design that have been learned from prior efforts
to prototype a policy engine are informative. Strauss [19] lists some useful design
requirements for a policy management engine, which are adapted to the cognitive
radio application and listed here without further comment.3
1. A policy condition must allow read access to variable attributes in a way that
the policy action can reference those attributes that matched the condition and
the attributes of the event that triggered the rule. Thus the run-time system
must bind variables to instances when passed to the condition and action.
2. There must be a construct to specify the value space in which the free variables
of conditions are evaluated that span all instances within a certain table or all
instances of a class.
3. A class that models a certain object must support accessory methods that allow
a policy to retrieve and manipulate an element and support computations. This
is not a requirement for the policy engine architecture itself but for the design
of interface.
4. Three types of time events may be required: Periodic events that trigger continuously for a given period, calendar events that trigger periodically at points in
time specified by calendar-type attributes, and one-shot events that trigger
exactly at a point in time specified by calendar-type attributes.
5. Another type of event is based on the reception of external notification, like such
as state reports, and these must be mapped to specific events. Details of the initiating notifications should be accessible through accessory methods of the events.
6. The policy run-time engine must support a mechanism to report errors and
optional tracing/debugging information so that users can monitor the policy
engine and the authors of policies can test and debug their policy codes.
7. The access to variables in conditions and actions may fail, and the policy runtime engine must be able to handle these situations in a way that policy code
can catch the error conditions and bring the variable to a determined state.
3
Any errors in adaptation are the responsibility of this author.
208
Cognitive Policy Engines
8. It must be possible to store and execute multiple policies independently. Their
codes must not share any name space.
9. A security mechanism is required to differentiate which users have access to
which roles on which policies (build on the existing SCA security component).
10. The notation of policies should remain declarative as much as possible, but
the programmatic policy system has to implement complicated actions with
code instead of just the goal.
11. It should be possible to avoid redundancy in a way that policies or policy
groups sharing rules and rules-sharing conditions or actions can be built by
referring common code instead of copying code fragments to increase
reusability and avoid some errors.
12. The policy engine should be able to pass arguments to policies when they are
activated by conditions based on messages or shared memory to reduce
redundancy.
6.6.3 Integration of the Radio into a Network Policy Management Architecture
Once PDP and PEP functionalities are accommodated in the policy engine, the
first question concerns what can be gained by embedding the cognitive radio in
networked management architecture. The discussion of the policy service functionality makes it clear that, at a minimum, the network policy architecture needs
to be concerned with policy dissemination. As Stone et al. [18] put it, “Communication is needed to and from the policy repository…. In many proposals the policy
repository is a directory, and therefore the appropriate access protocol would be
the Lightweight Directory Access Protocol.”
Of course, the actual “minimum” must do a little bit more. There must be
some kind of secure authorization to download policies to the radio. One way to
handle this is to architect the cognitive radio in such a manner that the only way
to update the policy database is to physically wire a connection to a special interface, thus relying on physical security to ensure the integrity of the policy database. This makes it very tedious and inconvenient to update the policy database,
however, particularly if a large number of radios are involved.
Since the purpose of the radio is to communicate, it makes sense to rely on network communications for policy updates. Sheridan-Smith [40] points out that peerto-peer management communications also eliminates problems with synchronization
of policy repositories because the radios can be informed of changes. Whenever
radios enter a new network, part of the session initiation can include setting up a policy management interface to synchronize policy databases as required.
209
Chapter 6
Naturally, no one would permit an unknown radio to remotely update the policy database without some kind of exchange of credentials to resolve trust issues.
Li et al. [41] introduce a family of role-based trust management languages for
representing policies and credentials. They define a “decentralized collaborative
system” being formed by several autonomous organizations desiring to cooperate
and share resources for their mutual benefit. Sheridan-Smith [40] also studies synchronization techniques for distributing the responsibility for policy-based network management (PBNM) across a set of independent autonomous PDPs: “The
use of the pull model is more efficient than a push model. Alternatively, the network management system is required to poll each device in turn to ensure that
they are operating correctly.” The push model is particularly problematic for the
cognitive radio because it will often be turned off and restarted or reinitialized.
Damianou et al. [23] introduce the next level of complexity: “It should be possible to dynamically update the policy rules interpreted by distributed entities to
modify their behavior.” When the concept of roles is introduced into the policy
language, the following premise is tacitly accepted: “It is not practical to specify
policies relating to individual entities—instead it must be possible to specify policies relating to groups of entities and also to nested groups such as sections within
departments, sites within organizations and within different countries” [23].
A simple view of policy in regard to cognitive radio networks is that network
policies constrain network communications. Specifically, network policy defines
the relationship between clients (users, applications or services) using spectrum
resources and the radios that provide access to those resources. (See Chapter 9 for
additional details on how the policy engine may manage network protocols.) Lee
et al. [33] take the argument to the next step: “Ideally, the management system
will use feedback from monitoring and network measurement to influence its
decisions about how to control the network configurations to improve efficiency,
manage service quality, or to deal with changes in the network environment.”
Thus a network policy management should incorporate a policy engine operating on network policies, receiving status updates from the policy service on individual cognitive radios, and downloading policy instructions to the members of
the network. Now the policy management architecture is self-similar, and, as
Boutaba and Znaty [31] state: “The whole network policy management system is
then logically constructed in a hierarchical domain structure where low level
domains provide their services to those of the upper layers. Domains are managed
according to a set of harmonized policies.”
At the network level, the policy manager must be focused on managing services by creating policies for individual devices based on the roles they will play in
210
Cognitive Policy Engines
the network architecture. This is the root of the distinction between network-level
policies and device-level policies. Sheridan-Smith [40] comments: “In general the
network policies can apply to more entities and are easier to read, write and comprehend, precisely because they are not specific about how the goal should be
achieved. Device-level policies can be specific about what needs to be done, but
they are difficult to read and complex to write and can apply only to a subset of
entities in the system.”
This, then, is the state of the art in automated policy management of complex networks. The difficulty involves how to automatically refine higher-level policies into
more specific management policies for domain members. This process involves planning for network operations and deriving specific policies for all the network elements to satisfy network goals. Stone et al. [18] define the research challenge as:
“refining the goals, partitioning the targets the policies affect or delegating responsibility to another manager who can perform this derivation. The main motivation for
understanding hierarchical relationships between policies is to determine what is
required for the satisfaction of policies. If a high-level policy is defined or changed, it
should be possible to decide which lower-level policies must be created or changed.”
6.7 The Future of Cognitive Policy Management
Having examined functionality and designs for cognitive policy management, this
section returns to the fundamental question “Why bother?” Damianou et al. [23]
reply that the “motivation for policy-based services is to support dynamic adaptability of behavior by changing policy without recoding or stopping the system.”
This chapter has already indicated how this is accomplished. This section considers promising military and commercial applications of policy management technology and examines the challenges.
6.7.1 Military Opportunities for Cognitive Policy Management
Military network communications are managed by a hierarchy of network operations centers that fit naturally into a recursive management model. In fact, the
military is unique in that it can hierarchically manage both the network resources
and the demand by users for resources, unlike commercial market applications in
which the financial goals always involve increasing user demand for services.
Figure 6.6 depicts the policy management hierarchy for communications
resources. This hierarchy also handles application performance, but a parallel
command structure exists for military personnel (users). Higher-level network
211
Chapter 6
Figure 6.6: Military policy
management hierarchy.
Task
Report
Planning
Networ
Network
Policies/
Plans
Policy Refinement
Operations
Center
Task
Task
Report
Tactical Ops Center
Tactical Ops Center
Planning
Planning
Policies/
Plans
Policy Refinement
Monitor/
Control
Unit Comms
Report
Coordinate/
Deconflict
Monitor/
Control
Unit Comms
Policies/
Plans
Policy Refinement
Monitor/
Control
Unit Comms
domains delegate communication management tasks to their subordinates to be
planned and executed on lower-level network resources. In this type of architecture, coordination is necessary, as Boutaba and Znaty [31] note: “Peer-to-peer
interactions take place in case of overlapping between domains to optimize plans
and deconflict policies.”
To handle the complexities of policy refinement that occur in this hierarchical
architecture, “the concept of management policy is introduced as an intermediate step
between goals and plans” [31]. In this case, “plans” refer to complex coordinated
actions, so the refinement process becomes: (1) lower-level policies are derived from
goals and missions passed down the hierarchy as tasks, (2) officers define action
plans consistent with the policies, and (3) subordinates are tasked to the plans.
The fact that DARPA has invested significant funds to support the development of various policy management applications for the military suggests that this
technology will filter down to the operational level. As network communications
systems such as the tactical Warfighter’s Internet (WIN-T) and the JTRS cluster
procurements come online, policy management should play an increasing role.
6.7.2 Commercial Opportunities for Spectrum Management
There is wide recognition that the FCC’s spectrum licensing and leasing practices
have become highly inefficient and have led to artificial shortages due to the
dedication of frequency bands for underutilized services. Due to the significant
212
Cognitive Policy Engines
financial investments in this area, progress will probably depend primarily on the
perception of financial incentives for the current spectrum lessees.
In particular, there has been some discussion of the possibility of spectrum
“micro-leases” that would permit occasional opportunistic use of certain spectral
bands in specific locations. Chiang et al. [42] describe protocol primitives to support a “Vickery auction” as a rational allocation of resources in a distributed system
[43]. The concept requires creation of a “broker space” and an “auction space,” by
which buyers and sellers could automatically purchase portions of the time/space/
frequency spectrum. In this protocol, the necessary steps are:
1. A selling agent informs the market of its intent to sell spectrum by issuing an
offer message.
2. Brokers link buyers to market opportunities.
3. Buyers submit bid messages to the seller.
4. The selling agent chooses its preferred buyer.
5. The transaction is settled (with a message) and the lease and moneys are conveyed between buyer and seller.
In some manner, it would be necessary to tie the policy management system
into this market to enable and disable the use of spectrum resources.
6.7.3 Obstacles to Adoption of Policy Management Architectures
This section summarizes a few of the challenges to widespread adoption of policy
management techniques for cognitive radios. Concerns about potential difficulty
in getting FCC certification for a radio managed by a cognitive policy engine have
already been addressed and will not be further discussed here.
It is one thing to design a policy management architecture and another thing
to demonstrate that its functionality is worth the effort to implement it. Lee et al.
[33] propose the design of experiments to test hypotheses and evaluate whether
the actions specified actually improve or worsen system performance: “the evaluation should proceed by examining a particular approach and making quantitative
measurements according to a well-defined methodology.” The DARPA PolySurv
program made some efforts in this direction, but the published literature consists
only of marketing brochures and advertisements from vendors.
Verma [44] recalls the need for policy-transformation logic that translates highlevel network administration policies to device-specific policies that can be automatically disseminated. The approach also ensures that high-level policies are
213
Chapter 6
mutually consistent, correct, and feasible. “The heart of policy management lies
in the policy translation logic as to how the policies will be represented and how
they will be managed” [44].
Boutaba and Znaty [31] see that “Today’s challenge is to provide automated
support not only for detecting and responding to trivial network and system
events, but also for the process of planning and policy making to handle more
complex situations.” They enlarge the scope of the automation problem: “Peer
negotiations for conflict avoidance should be handled as part of the proactive
management process, whereas negotiations for conflict resolution should be handled as part of the reactive management process. Both approaches should involve
all derivation of goals, policies, plans, and actions.”
Sheridan-Smith [40] is also concerned: “If multiple autonomous management
entities are making independent desiccations then each participant in the network
might behave in a manner that is not aligned with other parts of the network, or
other functions of the network which are independently managed. Coordination
will ensure that the behavior that is stipulated in the policies is correctly met by all
of the individual management entities.”
The perspective put forth in this chapter is that the adoption of a policy-based
spectrum management approach for cognitive radios will require standardization of
policy languages and dissemination techniques designed specifically for cognitive
radios and driven by market concerns. It seems likely that a cognitive policy engine
can be designed to function usefully in the context of cognitive radios. The real
question is probably not the technical feasibility of automatic policy management
architectures. Instead, it is whether market forces will be favorable to the adoption
of this technology and lead to integration across multiple platforms and vendors.
6.8 Summary
The focus of this chapter has been on the design of a practical cognitive policy
engine that would enable the cognitive radio to operate reasonably within a suitable policy space. The approach has been to explore the engineering requirements
for the policy engine in the context of a radio policy management domain that
might reasonably reflect the operational environment for a cognitive radio. The
scholarly literature shows what enabling technology is available and necessary to
realize this objective in the near term.
The basic syntactic and semantic mechanisms for representing spectrum and
network management policies were found to already be in existence in several
forms. Although there are limits on the semantic capabilities of extant policy
214
Cognitive Policy Engines
engines and languages, it is difficult to argue convincingly that there exists a significant technology gap in this area that would prevent cognitive radios from
being deployed in a fruitful manner.
If anything, the opposite conclusion is much more compelling. There is no
demonstrated requirement that calls for the representation of highly complex logical constraints on the operation of the cognitive radio. In fact, there is every reason to believe that implementations for policy engines in the first few generations
of cognitive radios will be more than capable of enforcing restrictions on policy to
constrain and optimize the behavior of the radio. It seems highly unlikely that
communication regulations will call upon the radio to dynamically resolve complex and sometimes inconsistent policies.
This chapter has reviewed the application of policy engines in commercial
networking environments and has found that the most difficult challenge lies not
in processing or enforcing the policies, but rather in coherently, unambiguously
specifying what behavior is desired. For this reason, existing policy-enabled network systems are generally focused on well-defined access control or network
configuration and performance optimization problems for which the desired
behaviors can be defined and agreed upon by users of the system.
The critical engineering problems that remain to be resolved for implementing
cognitive policy engines are primarily software design questions. Almost all of
these questions will be answered in terms that are specific to the particular kinds
of real-time software platforms and policy instantiations that have already been
addressed in this chapter. The interesting academic questions addressed primarily
relate to theoretical limitations of policy languages and abstract processing architectures that are much more complex than what will be initially embodied in the
cognitive radio. They are worthy of note, but should not be construed as obstacles
for implementing a cognitive policy engine. The real challenge lies in unambiguously specifying what behaviors are required of the cognitive radio, and achieving
consensus that such behaviors are sufficient for licensing the operation of the radio.
References
[1] J. Bemmel, P. Costa and I. Widya, “Paradigm: Event-Driven Computing,” Freeband
Awareness Deliverable D2.7a, October 2004; http://awareness.freeband.nl
[2] R. Chadha, G. Lapiotis and S. Wright, “Policy-Based Networking,” IEEE Network,
Vol. 16, No. 2, March 2002, pp. 8–9.
[3] K. Carey, K. Feeney and D. Lewis, “State of the Art: Policy Techniques for Adaptive
Management of Smart Spaces,” State of the Art Surveys, Release 2, Trinity College
Dublin, May 2003, pp. 58–66.
215
Chapter 6
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[26]
www.ntia.doc.gov/osmhome/allochrt.html
http://www.itu.int/ITU-R/
http://www.fcc.gov/sptf/reports.html
L. Kagal, T. Finin and A. Joshi, “A Policy Language for a Pervasive Computing
Environment,” in Proceedings of the IEEE 4th International Workshop Policies for
Distributed Systems and Networks, June 2003, pp. 63–77.
http://www-dse.doc.ic.ac.uk/research/policies/
http://www.ietf.org/html.charters/policy-charter.html
http://www.darpa.mil/ato/programs/XG/index.htm
http://www.darpa.mil/ato/programs/dynamiccoal.htm
C. Phillips, S. Demurjian and T. Tang, “Towards Information Assurance for Dynamic
Coalitions,” in Proceedings of the 2002 IEEE Workshop Information Assurance, June
2002.
A. Uszok, et al., “KAoS Policy Management for Semantic Web Services,” IEEE
Intelligent Systems, Vol. 19, No. 4, July/August 2004, pp. 32–41.
http://www.swsi.org
http://www.daml.org
http://www.w3.org
A. Uszok, et al., “KAoS Policy and Domain Services: Toward a Description-Logic
Approach to Policy Representation, Deconfliction, and Enforcement,” in
Proceedings of the IEEE 4th International Workshop Policies for Distributed
Systems and Networks, June 2003, p. 93.
G. Stone, B. Lundy and G. Xie, “Network Policy Languages: A Survey and a New
Approach,” IEEE Network, Vol. 15, No. 1, January 2001, pp. 10–21.
F. Strauss, Java Policy Management System: Design and Implementation Report,
Computer Science Department, Technical University Braunshweig, September 2001.
http://www.ibr.cs.tu-bs.de/projects/jasmin/
http://www-dse.doc.ic.ac.uk /research/policies
N. Dulay, et al., “A Policy Deployment Model for the Ponder Language,” in
Proceedings of the IEEE/IFIP International Symposium on Integrated Network
Management (IM’2001), May 2001.
N. Damianou, et al., A Survey of Policy Specification Approaches, Imperial College
of Science Technology and Medicine, April 2002 (submitted for publication).
N. Damianou, et al., “The Ponder Policy Specification Language,” in Proceedings
of the Policy 2001: Workshop Policies for Distributed Systems and Networks,
Springer-Verlag, LNCS, Vol. 1995, 2001, p. 18.
G. Tonti, et al., “Sematic Web Languages for Policy Representation and Reasoning:
A Comparison of KAoS, Rei, and Ponder,” in Proceedings of the 2nd International
Semantic Web Conference (ISWC 2003), Springer-Verlag, October 2003.
www.nortel.com
216
Cognitive Policy Engines
[27] B. Moore, et al., “Policy Core Information Model—Version 1 Specification,” IETF
RFC 3060, February 2001; http://www.rfc-editor.org/rfc/rfc3060.txt
[28] Y. Snir, et al., “Policy Quality of Service (QoS) Information Model,” IETF RFC 3644,
November 2003; http://www.rfc-editor.org/rfc/rfc3644.txt
[29] K. Feeney, D. Lewis and V.P. Wade, “Policy Based Management for Internet
Communities,” in Proceedings of the 5th IEEE International Workshop Policies for
Distributed Systems and Networks (POLICY’04), June 2004, p. 23.
[30] A. Uszok, et al., “Policy and Contract Management for Semantic Web Services,” in
Proceedings of the AAAI Spring Symposium on Semantic Web Services, March 2004.
[31] R. Boutaba and S. Znaty, “An Architectural Approach for Integrated Network and
Systems Management,” ACM-SIGCOM Computer Communication Review, Vol. 25,
No. 5, October 1995, pp. 13–39.
[32] A. Uszok, et al., “Applying KAoS Services to Ensure Policy Compliance for
Semantic Web Services Workflow Composition and Enactment,” in Proceedings of
the 3rd International Semantic Web Conference (ISWC 2004), Springer-Verlag,
LNCS, Vol. 3298, November 2004, pp. 425–440.
[33] S. Lee, et al., “Managing the Enriched Experience Network—Learning-Outcome
Approach to the Experimental Design Life-Cycle,” in Proceedings of Australian
Telecommunications, Networks and Applications Conference (ATNAC’03),
December 2003.
[34] H. Khurana and V.D. Gligor, “A Model for Access Negotiations in Dynamic
Coalitions,” in Proceedings of the Enterprise Security Workshop, 13th IEEE
International Workshop Enabling Technologies: Infrastructures for Collaborative
Enterprises (WETICE), June 2004.
[35] J.W. Lloyd, Foundations of Logic Programming, 2nd Edition, Springer-Verlag,
1987.
[36] http://www.ksl.stanford.edu/software/JTP
[37] www.sdrforum.org
[38] Software Communications Architecture, Prepared by Joint Tactical Radio System
Joint Program Office, version 2.2.1, April 2004.
[39] R. Montanari, G. Tonti and C. Stefanelli, “Policy-Based Separation of Concerns for
Dynamic Code Mobility Management,” in Proceedings of the 27th Annual
International Computer Software and Applications Conference (COMPSAC’03),
IEEE Computer Society Press, November 2003.
[40] N. Sheridan-Smith, A Distributed Policy-Based Network Management (PBNM)
System for Enriched Experience Networks (EENs), Assessment of Proposed
Doctoral Research, University of Technology, Sydney, November 2003.
[41] N. Li, J.C. Mitchell and W.H. Winsborough, “Design of a Role-Based Trust
Management Framework,” in Proceedings of the 2002 IEEE Symposium on
Security and Privacy, IEEE Computer Society Press, May 2002.
217
Chapter 6
[42] F. Chiang, et al., “Autonomic Service Configuration for Heterogeneous Telecommunication MASs with Extended Role-Based GAIA and JADEx,” in
Proceedings of the 2005 International Conference on Service Systems and Service
Management (ICSSSM 2005), June 2005.
[43] http://teleholonics.eng.uts.edu.au
[44] D. Verma, “Simplifying Network Administration using Policy Based Management,”
IEEE Network, Vol. 16, No. 2, March 2002, pp. 20–26.
218
CHAPTER 7
Cognitive Techniques:
Physical and Link Layers
Thomas W. Rondeau and Charles W. Bostian
Bradley Department of Electrical and Computer Engineering
Virginia Tech, Blacksburg, VA, USA
7.1 Introduction
This chapter discusses the expectation of a fully functional cognitive radio,
including the cognitive decision-making process using case-based theory and
genetic algorithms (GAs), to solve the multi-objective optimization problem
posed by such a radio. The presentation has as its basis the cognitive engine developed at the Virginia Polytechnic Institute and State University (Virginia Tech)—
Center for Wireless Telecommunications (VT-CWT).
This chapter focuses on the intelligent cross-layer optimization of physical
(PHY) and link (or medium access control, MAC) layers. The reader is encouraged to think beyond these two layers and consider how other layers, particularly
network and transport, can be adapted and optimized by using the same techniques.
Section 7.2 defines optimization for a cognitive radio. A discussion of the cognitive radio as a mix of artificial intelligence (AI) and wireless communications
follows in Section 7.3. Section 7.4 addresses the PHY and MAC layers, and considers which measurable radio settings and specifications (“knobs”) and which
radio and channel performance measures (“meters”) fall into which layer. Section
7.5 introduces multi-objective decision-making (MODM) theory to analyze the
radio’s performance, and presents the analogy of GA to represent the methodology. In Section 7.6, the tiered algorithm structure of the cognition loop, based on
modeling, action, feedback, and knowledge representation, is explored in detail.
219
Chapter 7
Tailoring GA techniques for a cognitive radio, based on the simple GA discussed in these earlier sections, is addressed in Section 7.7 by looking at casebased reasoning (CBR) and case-based decision theory (CBDT). The need for a
higher level of intelligence is the topic of Section 7.8, and Section 7.9 shows how
the collection of techniques addressed in this chapter creates a cognitive engine
capable of controlling and adapting a cognitive radio. Section 7.10 then summarizes the processes and ideas brought forth in this chapter.
7.2 Optimizing PHY and Link Layers for Multiple-Objectives
Under Current Channel Conditions
The goal of a cognitive radio is to optimize its own performance and support its
user’s needs. But what does “optimize” mean? It is not a purely selfish adaptation
where the radio seeks to maximize its own consumption of resources.
Consider the “tragedy of the commons” metaphor as it is often applied to
wireless communications [1].1 If two pairs of nodes are communicating on different networks using transmissions that overlap in time and frequency, they will
interfere. The nodes observe the interference as a low signal-to-interference and
noise ratio (SINR), and the classic response is then to increase transmitter power
to obtain a corresponding increase in SINR. As the transmitter on one link
increases its power, the other link will experience a lower SINR and respond by
increasing its power. Each radio will respond in turn to maximize its SINR at its
intended receiver by increasing its own transmit power. Each transmitter will ultimately increase its power to the limitations of the hardware. At this point, either
both links will have low SINR and therefore poor performance, or the link with
the more powerful transmitter will completely drown out the other. This is obviously a poor solution. Even in the latter scenario, a second glance shows this to be
bad for all concerned because now each radio is transmitting much more power
than is required, raising power consumption, reducing battery life, and increasing
potential interference to other users. This scenario is regularly reenacted in the
2.4 GHz industrial, scientific, and medical (ISM) band in which Bluetooth and
IEEE 802.11 devices are constantly creating cross-interference. Here 802.11 has a
higher transmit power, but Bluetooth has a protocol to continually repeat packet
transmissions until a successful transmission occurs.
1
Hazlett’s article [1] does not give a highly technical overview of this concept, but rather a regulatory analysis of the spectrum issues in general.
220
Cognitive Techniques: Physical and Link Layers
In contrast, a radio capable of understanding its environment and making
intelligent adaptations will recognize the problem it encounters with a competing
link trying to use the same band. While observing the other link, the cognitive
radio will not be limited by a simplistic understanding that “low SINR means I
should increase my transmitter power.” Instead, it will try other solutions, such as
altering the modulation or channel coding in ways that will improve frame error
rate (FER) performance in the channel. Or it will seek a channel free of interference and change its center frequency, thus relieving both radios of the burden of
fighting for the spectrum.
In a heavily congested spectral environment, changing frequency might not
be an acceptable solution. This is why it is important to look at all the possible
adjustments to the PHY and MAC layers to improve performance. A situation
might arise when all possible frequencies are in use, or the bandwidth required is
not available free from interference. At such times, a mix of cooperative techniques could be used, perhaps involving quadrature amplitude (and phase) modulation (QAM), spread spectrum techniques, orthogonal frequency division
multiplexing (OFDM), clever timing mechanisms, smart antenna beam forming or
null forming, or other operations that will allow sharing. This chapter addresses
methods of how to find a local or global optimum for the current channel environment. A generalized solution is not usually the answer; for example, spreading
techniques might better share spectrum than narrowband techniques, but a spread
spectrum system has its limits and might unnecessarily waste system resources.
A dynamic combination of techniques is really required to best adapt the radio in
real time for the specific local problems at hand. A well-designed cognitive radio
understands situations and analyzes how to best adapt all available radio communications parameters to present conditions for optimum performance.
First, a definition of optimization is in order. Within this chapter, a radio is
optimized when it achieves a level of performance that satisfies its user’s needs
while minimizing its consumption of resources such as occupied bandwidth and
battery power. To apply this, we need to understand what needs the user has and
how the radio performance can be adapted to meet these needs. The bottom line,
in general, is that the radio not overoptimize its external performance (its performance as observed by other radios) because this will have a negative impact on its
internal performance (computational power, complexity, and available internal
resources).2 The next section clearly defines the cognitive radio. The succeeding
2
Other, more sophisticated considerations are beyond the scope of this chapter.
221
Chapter 7
sections then discuss what parameters can be altered and how a cognitive radio
can use them to optimize its performance.
7.3 Defining the Cognitive Radio
Cognitive radios merge AI and wireless communications. The field is highly multidisciplinary, mixing traditional communications and radio work from electrical
engineering while applying concepts from computer science. Interestingly, Claude
Shannon, one of the early giants of communication theory, spent some of his time
thinking about and discussing intelligent machines, specifically, chess-playing
machines. In an article published in 1950, he discusses how computers could be
made to intelligently play chess, but he also lays out reasons why this has implications in other areas, such as designing filters, equalizers, relays, and switching circuits, routing, translating languages, organizing military operations, and making
logical deductions [2].
The cognitive radio architecture envisioned in the discussion in this chapter is
shown in Figure 7.1. Here, the intelligent core of the cognitive radio exists in the
cognitive engine. The cognitive engine performs the modeling, learning, and optimization processes necessary to reconfigure the communication system, which
appears as the simplified open systems interconnection (OSI) stack [3]. The cognitive engine takes in information from the user domain, the radio domain, the
policy domain, and the radio itself. The user domain passes information relevant
Figure 7.1: Generic cognitive radio
architecture. This architecture has a
cognitive engine to observe behaviors
of the OSI protocol stack and propose
optimizations based on the current
environment. The policy engine
determines whether the hardware can
support those optimizations as well as
whether it is allowed to by regulatory
and network control.
User Domain
Operational Cognitive Radio Platform
Communication
System
Application
Cognitive
Engine
Transport
Network
Policy
Engine
Link/MAC
PHY
External Environment and RF Channel
Policy Domain
222
Cognitive Techniques: Physical and Link Layers
to the user’s application and networking needs to help direct the cognitive
engine’s optimization. The radio domain information consists of radio frequency
(RF) and environmental data that could affect system performance such as propagation or interference sources. The policy engine receives policy-related information from the policy domain. This information helps the cognitive radio decide
on allowable (and legal) solutions and blocks any solutions that break local
regulations.
Most of the topics shown in Figure 7.1 are covered in more detail in this chapter as their relationship to the cognitive engine are developed. The policy engine
and policy domain will be left to experts of that field, but are included here for
completeness.
7.4 Developing Radio Controls (Knobs) and Performance
Measures (Meters)
The first problem in dealing with cognition in a system is to understand (1) what
information the intelligent core must have and (2) how it can adapt. In radio, we
can think of the classical transmitters and receivers as having adjustable control
parameters (knobs) that control the radio’s operating parameters. Think of a frequency modulation (FM) broadcast receiver with a tuning knob to select which
station you are listening to as well as equalizer knobs to adjust the sound quality
to your liking. Radio performance metrics are referred to as meters. What follows
is an analysis of the knobs and meters important to a cognitive radio on the PHY
and link layers.
Huseyin Arslan has developed a useful layered classification of knobs and
meters (observable parameters and writable parameters in his notation), which is
summarized and expanded in Table 7.1.3
7.4.1 PHY- and Link-Layer Parameters
Knobs
The knobs of a radio are any of the parameters that affect link performance and
radio operation. Some of these are normally assumed to be design parameters, and
others are usually assumed to be under real-time control of either the operator or
the radio’s real-time control processes. Figure 7.2 shows a simple system diagram
of the PHY- and link-layer portions of a transmitter. In the PHY layer, center
3
Huseyin Arslan, personal communication.
223
Chapter 7
Table 7.1: Example tabulation of knobs and meters by layer.
Layer
Meters* (observable parameters)
Knobs (writable parameters)
NET
Packet delay
Packet jitter
CRC check
ARQ
FER
Data rate
Packet size
Packet rate
Source coding
Channel coding rate and type
Frame size and type
Interleaving details
Channel/slot/code allocation
Duplexing
Multiple access
Encryption
Transmitter power
Spreading type
Spreading code
Modulation type
Modulation index
Bandwidth
Pulse shaping
Symbol rate
Carrier frequency
Dynamic range
Equalization
Antenna beamshape
MAC
PHY
BER
SNR
SINR
RSSI
Path loss
Fading statistics
Doppler spread
Delay spread
Multipath profile
AOA
Noise power
Interference power
Peak-to-average power ratio
Error vector magnitude
Spectral efficiency
*AOA: angle of arrival; ARQ: automatic repeat request; CRC: cyclic redundancy check;
MAC: medium access control; NET: network layer; PHY: physical layer; RSSI: received signal
strength indicator; SINR: signal-to-interference and noise ratio; SNR: signal-to-noise ratio.
frequency, symbol rate, transmit power, modulation type and order, pulse-shape
filter (PSF) type and order, spread spectrum type, and spreading factor can all be
adjusted. On the link layer are variables that will improve network performance,
including the type and rate of the channel coding and interleaving, as well as
access control methods such as flow control, frame size, and the multiple access
technique.
224
Cognitive Techniques: Physical and Link Layers
Spread
Spectrum
Modulation
PSF
Physical Layer
LOIF
Interleaver
FEC
Encrypt
Power
Control
LORF
Source
Coding
Data
Link Layer
Framing
Flow Control
Duplexing
Link Access
Figure 7.2: Generic transmitter PHY and MAC layers. Many of the radio control parameters
(knobs) apply to these elements of the block diagram, resulting in profound impact on radio
performance (meters). (Note: FEC: forward error correction.)
Meters
Once we understand what knobs are available to optimize the radio system performance, we must understand how these changes affect the radio channel and system
performance to allow an autonomous, intelligent decision-maker to adapt the radios.
Performance is a measure of the system’s operation based on the meter readings. In optimization theory, the meters represent utility and cost functions that
must be maximized or minimized for optimum radio operation. All of these performance analysis functions constitute objective functions. In an ideal case, we can
find a single-objective function whose maximization or minimization corresponds
to the best settings. However, communication systems have complex requirements
that cannot be subsumed into a single-objective function, especially if the user or
network requirements change. Metrics of performance are as different for voice
communications as they are for data, e-mail, web browsing, or video conferencing.
The types of meters represent performance on different levels. On the PHY
layer, important performance measurements deal with bit fidelity. The most obvious meters are the signal-to-noise ratio, or a more complex SINR. The SINR has a
direct consequence on the bit error rate (BER), which has different meanings for
225
Chapter 7
different modulations and coding techniques, usually nominally determined by the
SINR ratio, Eb/(N0 I0), where Eb is energy per bit, N0 is noise power per bit,
and I0 is interference power per bit. On the link layer, the packet fidelity is an
important metric, specifically the packet error rate (PER).
There are more external metrics to consider as well, such as the occupied
bandwidth and spectrum efficiency (number of bits per hertz) and data rate. The
growth of complexity to optimize multiple metrics quickly becomes apparent.
Each metric has unique relationships with the other metrics, and different knobs
alter different metrics in different ways. For example, altering the modulation type
to a higher order will increase the data rate but worsen the BER.
Internal metrics also are involved in decision-making. To decrease the FER,
we could use a stronger code, but this increases the computational complexity of
the system, increasing both latency as well as the power required to perform the
more complex forward error correction (FEC) operation. Decreasing the symbol
rate or modulation order will decrease the FER as well without increasing the
demands of the system, but at the expense of the data rate. Figure 7.3 begins to
Figure 7.3: Directed graph indicating how one objective (source) affects another objective
(target).
226
Cognitive Techniques: Physical and Link Layers
expand upon these relationships, where the direction of the arrow indicates that
optimization in the source objective affects the target objective. Ongoing work
fully defining these relationships should lead to more knowledge for the adaptation and learning system to use.
7.4.2 Modeling Outcome as a Primary Objective
The basic process followed by a cognitive radio is that it adjusts its knobs to achieve
some desired (optimum) combination of meter readings. Rather than randomly trying
all possible combinations of knob settings and observing what happens, it makes
intelligent decisions about which settings to try and observes the results of these trials. Based on what it has learned from experience and on its own internal models of
channel behavior, it analyzes possible knob settings, predicts some optimum combination for trial, conducts the trial, observes the results, and compares the observed
results with its predictions, as summarized in the adaptation loop of Figure 7.4. If
results match predictions, the radio understands the situation correctly. If results do
not match predictions, the radio learns from its experience and tries something else.
Figure 7.4: Adaptation
loop.
Analyze
Possible
Knobs
Predict
Optimum
Settings
Observe and
Model
Conduct
Trials
Compare to
Predictions
Observe
Results
This operational concept employed for the cognitive radio resembles closely
some of the current thinking about how the human brain works [4]. The argument
holds that human intelligence is derived from predictive abilities of future actions
based on the currently observed environment. In other words, the brain first models the current situation as perceived from the sensor inputs, and it then makes a
prediction of the next possible states that it should observe. When the predictions
do not match reality, the brain does further processing to learn the deviation and
incorporate it with its future modeling techniques. Although knowledge of how
the human brain actually works is still uncertain, this predictive model is a good
227
Chapter 7
one to work from because it brings together the necessary behavior required from
the cognitive radio.
As an example of the mathematics involved in this process, consider observations of BER and SINR. BER formulas are generally represented by the complementary error function or the Q function (Eq. (7.1)) as a function of the SINR:
p
∫ et dt
2
x
1
Q( x ) Q( x ) 2
erfc( x ) 2p
∫ et
2
x 0
(7.1)
x
⎛ x⎞
1
erfc ⎜⎜ ⎟⎟⎟
⎜⎝ 2 ⎟⎠
2
where x is the SINR. A computationally efficient calculation for BER formulas uses
an approximation to the complimentary error function. Eq. (7.2) shows two approximations, one for small values of x (3) and one for large values of x (3) [5, 6].
Figure 7.5 compares the results of the formulas to the actual analytical function:
⎧⎪
3
5
⎛
⎞
⎪⎪1 1 ⎜⎜x x x ⎟⎟ for x 3 with 40 items in the series
⎟⎟
⎜⎜
⎪⎪
3
10
⎝
⎠
p
erfc( x ) ⎨ 2
(7.2)
⎪⎪ ex ⎛
⎞⎟
1
1.3
1⋅3⋅5
⎜
⎪⎪
⎜1 2 2 4 3 6 ⎟⎟ for x 3 with 10 items in the series
⎟⎠
2x
2 x
2 x
⎪⎩x p ⎜⎝
Figure 7.5: erfc approximation
Eq. (7.2)—compared to
analytical formula—Eq. (7.3).
1
Analytical
Approximation
0.9
0.8
erfc(x)
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
0.5
1
1.5
2
2.5
x
228
3
3.5
4
4.5
5
Cognitive Techniques: Physical and Link Layers
These formulas are useful because the normal approximation for the erfc function
is valid only for large x (x 3), and the Q function is too computationally intensive to calculate. Because a cognitive radio needs to perform a lot of these calculations, we need efficient equations; these equations trade accuracy for computational
time based on the number of terms included in the series expansion. Similar lines
of thought must go into developing each objective calculation.
The exact representation of the BER formula depends on the channel conditions and modulation being used. A standard BER formula for binary phase shift
keying (BPSK) signals in an additive white Gaussian noise (AWGN) channel is:
⎛
1
C ⎞⎟⎟
⎜⎜
Pe erfc ⎜ T0 B ⎟⎟
⎜⎝
2
N ⎟⎠
(7.3)
where T0 is the symbol period, B is the bandwidth, C is the signal carrier energy,
and N is the noise power.
In a fading channel with a probability density function (PDF) of p(x), the BER
of a signal is defined as:
Pe ∫ PAWGN ( x ) p( x ) dx
(7.4)
0
where PAWGN(x) is the BER formula in an AWGN channel.
The radio observes the BER and SINR value. If these are consistent according
to the above formulas, the radio can assume that the channel is behaving predictably. It can then turn knobs that directly affect SINR, for example starting
with the easiest, transmitter power.4 If the transmitter power is already at the
allowable limit, the radio may lower the data rate to change the occupied bandwidth and therefore increase the average energy per bit. If the BER and SINR are
not consistent with the known formulas, the radio might assume, for example, that
4
The difference between predicted link performance and actual link performance includes both
errors in the estimate of propagation channel losses, and nonlinear effects arising from interference
and multipath. Performance on an AWGN channel in the absence of multipath is predictable.
When this performance difference is significantly large, it may become clear that transmit power
alone is inadequate to achieve the necessary performance. Thus, the performance difference is a
good indicator of the need to invoke these cognitive radio techniques to optimize performance in
the presence of unusual channel behavior.
229
Chapter 7
the channel is dispersive and opt to change the carrier frequency rather than the
transmitter power.
This analysis has dealt with only a single objective. The radio can, in fact,
read a number of meters, and each of these can be some objective function we
may wish to optimize. Standard communications theory can lead us to the methods of mathematically modeling each objective [7]. The communications analysis
tools are fairly standard, and another aspect is the consideration for how to efficiently realize each objective function. For each, we must carefully choose the
proper analytical expression that is not too computationally complex.
7.5 MODM Theory and Its Application to Cognitive Radio
The wireless optimization concept has already been described through an analysis
of the many objective functions (dimensions) of inputs (knobs) and outputs
(meters). In this scenario, the interdependence of the objectives to each other and
to various knobs makes it difficult to analyze the system in terms of any one single
objective. Furthermore, the needs of the user and of the network cannot all be met
simultaneously, and these needs can change dramatically with time or between
applications. For different users and applications, radio performance and optimum
service have different meanings. As a simple example, e-mail has a much different
performance requirement than voice communications, and a single-objective function would not adequately represent these differing needs.
Without a single-objective function measurement, we cannot look to classic
optimization theory for a method to adapt the radio knobs. Instead, we can analyze the performance using MODM criteria. MODM theory allows us to optimize
in as many dimensions as we have objective functions to model.
MODM work originated about 40 years ago and has application in numerous
decision problems from public policy to everyday decisions (e.g., people often
decide where to eat based on criteria of cost, time, value, customer experience,
and quality). An excellent introduction to MODM theory is given in a lecture
from a workshop held on the subject in 1984 [8]. Schaffer then applied MODM
theory to create a GA capable of multi-objective analysis in his doctoral dissertation [9]. Since then, GAs have been widely used for MODM problem-solving.
GAs are addressed in detail in Section 7.5.5.
7.5.1 Definition of MODM and Its Basic Formulation
At their core, MODMs are a mathematical method for choosing the set of
parameters that best optimizes the set of objective functions. Eq. (7.5) is a basic
230
Cognitive Techniques: Physical and Link Layers
representation of a MODM method [10]:
min/max{y} f ( x ) [ f1 ( x ), f2 ( x ), … , fn ( x )]
subject to: x ( x1, x2 , … , xm ) ∈ X
y ( y1, y2 , … , yn ) ∈ Y
(7.5)
Here all objective functions are defined to either minimize or maximize y, depending on the application. The x values (i.e., x1, x2, etc.) represent inputs and the y
values represent outputs. The equation provides the basic formulation without prescribing any method for optimizing the system. Some set of objective functions
combined in some way will produce the optimized output. There are many ways
of performing the optimization. Section 7.6 discusses one of the more complex
but useful methods of solving MODM problems for cognitive radios.
7.5.2 Constraint Modeling
An added benefit of MODM theory implicit in its definition is the concept of constraints. The inputs, x, are constrained to belong to the allowed set of input conditions X, and all output must belong to the allowed set Y. This is important for
building in limitations for hardware as well as setting regulatory bounds.
7.5.3 The Pareto-Optimal Front: Finding the Nondominated Solutions
In an MODM problem space, a set of solutions optimizes the overall system, if
there is no one solution that exhibits a best performance in all dimensions. This
set, the set of nondominated solutions, lies on the Pareto-optimal front (hereafter
called the Pareto front). All other solutions not on the Pareto front are considered
dominated, suboptimal, or locally optimal. Solutions are nondominated when
improvement in any objective comes only at the expense of at least one other
objective [11].
The most important concept in understanding the Pareto front is that almost all
solutions will be compromises. There are few real multi-objective problems for
which a solution can fully optimize all objectives at the same time. This concept
has been referred to as the utopian point [12]; this point is not considered further
here because in radio modeling problems only very rarely do situations have a
utopian point. One only has to consider the most basic radio optimization problem
to see this: simultaneously minimize BER and power. Figure 7.6 shows the ideal
BER curve of a BPSK signal. Here, point A is a nondominated point that minimizes power at the expense of the BER, and moving down the curve to point B,
231
Chapter 7
which is also nondominated, optimizes for BER at the expense of greater power
consumption. Point C is a dominated point that represents a suboptimal solution
of using differential phase shift keying (DPSK) to minimize the complexity of
carrier phase tracking in high multipath mobile applications. The MODM problem then reduces to a trade-off decision between low BER, low power consumption, and complexity due to other system constraints.
BPSK
DPSK
A
2
4
C
BER
Figure 7.6: Pareto front of a BPSK
BER curve compared to a
dominated solution of DPSK.
Condition A is least power,
condition B is lowest BER,
and condition C is least complexity.
6
8
B
10
0
5
10
15
20
Eb/N0 (dB)
7.5.4 Why the Radio Environment Is a MODM Problem
The primary objectives developed thus far have different meanings and importance, depending on the user’s needs and channel conditions. To optimize the
radio behavior for suitable communications, we must optimize over many or all of
the possible radio objectives. Take, as an example, the BER curve. In the twodimensional plot, the result of the using a differential receiver was suboptimal
because we were concerned with BER and power. But if we add complexity as an
objective, or the need for a solution that does not require phase synchronization,
such as might be necessary in highly complex and dynamic multipath, then using
a differential receiver for lower complexity becomes an important decision along
a third dimension. The resulting search space is N-dimensional and, due to the
complex interactions between objectives and knobs, the space is difficult to define
and certainly not linear, or even simply convex as is desirable in optimization theory [13]. These interactions are often difficult to characterize and predict, and so
we must analyze each objective independently and use MODM theory to find an
optimal aggregate set of parameters.
232
Cognitive Techniques: Physical and Link Layers
What further enhances the complexity of the search space is that it will change
depending on the user and the application. For certain users or applications, different objectives will mean different levels of quality. As the overall optimization
is to provide the best quality of service (QoS) to the user, there is no single search
space that can account for all the variations in needs and wants from a given radio.
From this analysis emerge a few important points about how to analyze the
multiple-objectives used in optimizing a radio:
●
●
●
●
●
Many objectives exist, creating a large N-dimensional search space.
Different objectives may be relevant for only certain applications/needs.
The needs and subjective performances for users and applications vary.
The external environmental conditions determine what objectives are valid and
how they are analyzed.
We may search for regions where multiple performance metrics meet acceptable performance, rather than searching for optimal performance.
This leads to a need for an MODM algorithm capable of robust, flexible, and
online adaptation and analysis of the radio behavior. The clearest method of realizing all the needs of the problem statement is the GA, which is widely considered
the best approach to MODM problem-solving [9, 10, 14–16]. Section 7.5.5 discusses the approach to GAs and Section 7.9.1 shows how to add user/application
flexibility into the algorithm.
7.5.5 GA Approach to the MODM
Analyzing the radio by using a GA is inspired by evolutionary biological techniques. If we treat the radio like a biological system, we can define it by using an
analogy to a chromosome, in which each gene of the chromosome corresponds to
some trait (knob) of the radio. We can then perform evolutionary-type techniques
to create populations of possible radio designs (waveform, protocols, and even
hardware designs) that produce offspring that are genetic combinations of the parents. In this analogy, we evolve the radio parameters much like biological evolution to improve the radio “species” through successive generations, with selection
based on performance guiding the evolution. The traits represented in the chromosome’s genes are the radio knobs, and evolution leads toward improvements in the
radio meters’ readings.
GAs are a class of search algorithms that rely on both directed searches
(exploitation) and random searches (exploration). The algorithms exploit the
233
Chapter 7
current generation of chromosomes by preserving good sets of genes through the
combination of parent chromosomes, so there is a similarity between the current
search space and the previous search space. If the genetic combination is from
two highly fit parents, it is likely that the offspring is also highly fit. The algorithms also allow exploration of the search space by mutating certain members of
the population that will form random chromosomes, giving them the ability to
break the boundaries of the parents’ traits and discover new methods and solutions. While providing the iterative solution through genetic combination, the
randomness helps the population escape a possible local optimum or find new
solutions never before seen or tried, even by a human operator. In effect, this last
quality provides the algorithm with creativity.
Introduction to GAs
GAs are often useful in large search spaces, which can enable their use in many
situations. A GA is a search technique inspired by biological and evolutionary
behavior. The GAs use a population of chromosomes that represent the search
space and determine their fitness by a certain criterion (fitness function). In each
generation (iteration of the algorithm), the most fit parents are chosen to create
offspring, which are created by crossing over portions of the parent chromosomes
and then possibly adding mutation to the offspring. The crossover of two parent
chromosomes tries to exploit the best practices of the previous generation to create a better offspring. The mutation allows the search algorithm to be “creative”:
that is, it can prevent the GA from getting stuck in a local maximum by randomly
introducing a mutation that may result in improved performance metrics possibly
closer to the global maximum, according to the optimization criteria.
To realize the GA, we follow the practices described by Goldberg [17]:
1. Initialize the population of chromosomes (radio/modem design choices)
2. Repeat until the stopping criterion
(a) Choose parent chromosomes
(b) Crossover parent chromosomes to create offspring
(c) Mutate offspring chromosomes
(d) Evaluate the fitness of the parent chromosomes
(e) Replace less fit parent chromosomes
3. Choose the best chromosome from the final generation
This process is illustrated in detail in the next section.
234
Cognitive Techniques: Physical and Link Layers
The Knapsack Example
To explain the operation of a simple GA, we examine the knapsack problem [18],
which is a classic nondeterministic polynomial time (NP) complete5 problem [19],
also called the subset-sum problem (SSP). The knapsack problem is defined by
the task of taking a set of items, each with a weight, and fitting as many of these
items into the knapsack while coming as close to, but not exceeding, the maximum weight the knapsack can hold. Mathematically, the knapsack problem is
shown by Eq. (7.6), where K is the maximum weight the knapsack can hold, and
Ns is the number of items in the set, S. The problem is represented by a weight
vector w and a vector x that is a vector of 1’s and 0’s that indicates whether an
item is present in the knapsack:
Ns
max ∑ xi wi
i 1
(7.6)
Ns
subject to: ∑ xi wi K
i 1
Following the practices just enumerated, use the following steps.
Step 1. Initialize Chromosomes
We introduce radio chromosome selection as a problem similar to knapsack selection. In this case, the problem consists of choosing the right set of items to place
in the knapsack, so the chromosome will represent the vector x and consist of 1’s
and 0’s as shown in Figure 7.7. Each gene is 1-bit wide and so very compact and
easy to manipulate mathematically.
Figure 7.7: Chromosome representation
of knapsack item vector.
x1
x2
x3
...
x Ns
Step 2a. Choose
The choice of the parent chromosomes determines how random the population will remain and how much memory the biological system retains of
5
NP is a class of decision problems whose positive solutions can be verified in polynomial time
given the right information. It is a description of the difficulty of a computational problem, where
NP-complete problems are the most difficult.
235
Chapter 7
its fitness for its environment. Like biological evolution, the most fit
parents are more likely to be chosen to produce offspring; however, it is still
possible to choose an unfit parent. The more random the selection process,
the more random the population will be in each generation. Conversely,
the more the decision is weighted toward the most fit parents, the faster the
convergence to local optima will occur. This trade-off is referred to as selection
pressure.
As in all GA properties, the choice of how parents are chosen comes down
to how random the population will remain and how fast convergence will occur.
Although convergence time is highly important, fast convergence may not always
be beneficial, depending on the shape of the fitness landscape, that is what the
solution space looks like. If the fitness landscape has many local maxima and the
GA is searching for the global maximum, fast convergence is more likely to lock
on to a local maximum, and take many generations and mutations to get beyond
the local maximum to find the global maximum. A more diverse population will
generate many more solutions that are more likely to find the global maximum.
Five different methods of selection were analyzed by De Jong in his doctoral
dissertation [20].
The method used in this GA is called tournament selection. During reproduction, this selection scheme chooses as many parents as there are members in the
population with replacement (i.e., the same parent may be chosen multiple times).
In the selection, two parents are randomly chosen and the more fit parent wins the
tournament and is selected for reproduction.
Step 2b. Crossover
Crossover is performed on two parents to form two new offspring. The GA has a
crossover probability that determines if crossover will happen. A randomly generated floating-point value is compared to the crossover probability, and if it is less
than the probability, crossover is performed; otherwise, the offspring are identical
to the parents. If crossover is to occur, one or more crossover points are generated,
which determines the position in the chromosomes where parents exchange genes.
In this GA, once two parents are selected, two crossover points are randomly
generated.
Figure 7.8 illustrates the crossover operation, but, for the simplicity of the
figure, the genes represent a knapsack problem with only eight items. The
genes after the first crossover point and before the second crossover point are
interchanged in the parents to form the new offspring.
236
Cognitive Techniques: Physical and Link Layers
Figure 7.8: Parent chromosomes
crossover. Crossover occurs at points
2 and 6 to create offspring chromosome
sets.
Parent 1:
x10
x11
x12
x13
x14
x15
x16
x17
Parent 2:
x20
x21
x22
x23
x24
x25
x26
x27
Offspring 1:
x20
x21
x12
x13
x14
x15
x26
x27
Offspring 2:
x10
x11
x22
x23
x24
x25
x16
x17
Step 2c. Mutate
After the offspring are generated from the selection and crossover, the offspring
chromosomes may be mutated. Like crossover, there is a mutation probability. If a
randomly selected floating-point value is less than the mutation probability, mutation is performed on the offspring; otherwise, no mutation occurs. Mutation is
performed by randomly selecting a gene in the offspring’s chromosome and generating a new value based on some PDF (often a uniform or Gaussian distribution). In the knapsack problem, a gene may be reset to either a 1 or 0 at random
(other techniques invert the gene).
Steps 2d and 2e. Evaluate and Replace
Evaluation (step 2d) is done on the multi-objective performance of the offspring
from the most recent genetic cycle to determine whether to continue the genetic
process (step 2e). Evaluation is probably the most important piece of the GA aside
from the initial chromosome definition. Choice of the fitness evaluation is vital to
convergence and proper application.6 The best set of offspring from among all
those evaluated are ordered by performance and added to the rank-ordered list of
chromosomes—say, for example, the top 10 or the top 100. Evaluation determines
whether the gene pool is improving with the previous cycles—say, the last five
6
For the knapsack problem, we try to find the chromosome with fitness defined by Eq. (7.6). If the
constraint condition is not met, the fitness is set to 0.
237
Chapter 7
cycles. If the performance is sufficiently above the goal behavior, the iterative
process may terminate and the procedure goes to step 3. If, however, the performance over the previous cycles continues to improve but the goal performance is
not yet met, the algorithm iterates back to step 2a. If the performance is unable to
improve in previous cycles, the mutation rate may be adapted to a higher mutation
rate. The evaluation function is very problem-specific, and is one of the main
sources of research in the multi-objective GA to optimize cognitive radios.
Step 3. Results: Choose Best Chromosomes
To analyze the GA performance, we often graphically represent the performance
over many generations, such as the knapsack example shown in Figure 7.9. This
figure shows the general trends of a GA as a maximization problem with steady
increase in the fitness over generations. In this example, we set the crossover
probability at 0.95 and mutation probability at 0.05. There are 20 members of the
population, and in each generation 17 members of the population are replaced by
offspring. The problem contains 200 items, and each item could take on any
floating-point weight value between 0 and 20. The maximum weight the knapsack
can hold is a randomly generated integer of 810. The trends shown in Figure 7.9
show steady improvement toward the goal of 810, coming very close by the 155th
generation with a value of 809.87.
850
800
Fitness
750
700
650
600
Best Fitness
550
Average Fitness
500
Worst Fitness
450
0
20
40
60
80
100
120
140
160
Generation
Figure 7.9: Performance graph of knapsack GA. The fitness metric improves with each
generation, nearly reaching the maximum value after 155 generations.
238
Cognitive Techniques: Physical and Link Layers
As shown in Figure 7.9, a GA performs by making great initial progress, but
then slows down as it approaches the optimal value, and even then, it does not
perfectly achieve the optimum. Section 7.5 explores the use of GAs in radio optimization problems. Radio optimization must be real time, but it does not have
to completely optimize under real-time constraints. Instead, we want to provide
better responses as opposed to a best response for the immediate future. GAs
quickly improve their performances in the first few generations; by the time
the knapsack problem reaches its first plateau around generation 30, we have
a useful solution. Looking at what we need out of the performance of a radio
instead of just what we want, the GA can give us very usable solutions very
quickly. As time and resources permit, the final iterations of the GA can find
increasingly better solutions; other techniques, introduced in Section 7.5, such
as CBDT, adaptive GAs, and distributed computing, can improve the solutions
even faster.
Another point to consider involves the temporal features of optimum solutions. In the future, situations envisioned for the cognitive radio, the environments, both propagation and interference, as well as the user’s and application’s
needs, will change, and with them the optimum solution. We therefore want a system that will continue to change and adapt with the needs of the situation for continuous improvement in performance.
7.6 The Multi-objective GA for Cognitive Radios
7.6.1 Cognition Loop
The primary goal of the cognitive engine is to optimize the radio, and the secondary functions are to observe and learn in order to provide the knowledge required
to perform the adaptation. A cognitive radio becomes a learning machine through
a tiered algorithm structure based on modeling, action, feedback, and knowledge
representation, as shown in the cognition loop of Figure 7.10.
This section presents the high-level structure of the Virginia Tech cognitive
radio solution as well as the background of the basic theories and procedures that
were followed in its development. The following sections discuss the system in
detail.
Modeling
Modeling means that the machine must have some representation of the outside
world to which it can respond. Models represent external influences, such as the
239
Chapter 7
Scenario
Synthesizing
Environment
Observation
Case Identified
Link Condition
User/Policy
Knowledge Base
Case-Based
Decision
Making
Reasoning
Radio Hardware
Success Memorized
Radio
Hardware
Case Report
Apply Experience
Bad Trail Overwritten
Strategy Instruction
Link Configure
Optimization
Performance
Estimation
WSGA
Initialization
Objectives
Constraints
Figure 7.10: Cognition loop. Note the outer loop to observe and adapt a radio and the
inner loop for learning.
user’s network needs and actions, the radio spectrum and propagation environment, and the governing regulatory policy. As each of these influences changes,
the radio must adjust itself to satisfy the new operating environment.
A primary function of the modeling system is to provide environmental context to the cognitive radio, such as propagation effects and the presence of other
radios, which might be either cooperative or potential interferers.
Another goal of the modeling system is to monitor the data sent by the user
and use it to determine the QoS parameters the cognitive radio must provide. This
is a learning domain concept that a neural network could be employed to help
solve, and the research group at VT is currently investigating the possibilities. The
regulatory policy modeling and interpretation will come from other efforts, such
as the Defense Advanced Research Projects Agency (DARPA) NeXt Generation
(XG) project [21].
Figure 7.11 provides background about how Virginia Tech’s work fits into the
literature of machine learning and AI. It shows the VT-CWT cognitive engine in
more detail, emphasizing its three major subsystems.
The cognitive system module (CSM) is responsible for learning and the wireless system genetic algorithm (WSGA) handles the behavioral adaptation of the
radio, based on what it is told to do by the CSM. The modeling system observes
the environment from many different angles to develop a complete picture. The
CSM holds two main learning blocks: the evolver and the decision maker, which
takes feedback from the radio that allows the evolver to properly update the
knowledge base to respond to and direct system behavior.
240
Cognitive Techniques: Physical and Link Layers
The VT cognitive
engine can
perform this
function if
necessary
The channel can be RF
or optical using any
modulation format
Channel
(Data/RF)
Channel Probe
Radio Hardware
Radio Baseband Processor
(PHY and MAC Layer)
The API defines
the interface to any
specific platform
API-Engine to Radio Interface
User Domain
Security
Policy
Domain
How to play:
use channel and
user models and
radio parameters
to enable radio
evolution
Cognitive Engine
Modeling System
– User, Channel,
Regulations
Channel, User
and Regulatory
Model
Resource
Monitor
Wireless System
Genetic Algorithm
(WSGA)
System
Chromosome
and Simulated
Fitness Values
Radio
Parameters/
Performance
Statistics
The radio can be a legacy
system or the latest
software-defined radio
Evaluation
Functions, Child
Chromosomes,
Regulations
CSM
(Learning Machine, Evolutionary Algorithm, Memory)
Take action
to adapt radio
The VT cognitive engine
is hardware independent
and can make any radio
with “knobs” and
“meters” cognitive
Figure 7.11: The VT cognitive engine. This hardware-independent modeling, learning,
and adaptation system interfaces to radio hardware via a radio-specific application
programmable interface.
Action
Actions are taken by the WSGA in creating a new radio configuration. This system is instructed by the CSM when a new model is observed and change is
required in the radio operation. The WSGA creates a set of parameters that best
optimize the radio for the new settings.
The WSGA treats the radios like biological creatures, where genes represent
such radio traits as power, frequency, modulation, FEC coding, filtering, spreading, and so on. Groups of genes constitute chromosomes. The WSGA applies a
multi-objective GA to a population of chromosomes and evaluates the results by
using the objective functions and weights provided by the knowledge base. The
functions determine which outputs are important, such as the BER, PER, power
consumption, interference potential, and so forth, and the weights determine the
relative importance of each function. After running the WSGA through some
number of generations (terminating after convergence has been observed or following a predetermined number of generations, whichever occurs first), the chromosome that performs best in the largest number of objectives is chosen as the
new set of radio parameters. Throughout this process, the chromosomes will be
241
Chapter 7
analyzed continually with respect to validation routines to ensure no combination
is chosen for the final radio setting that would violate regulatory authority rules
or the network operator’s operational procedures. Detailed information about the
WSGA and experimental results have been published [22], and a summary is
provided in an example in Section 7.6.5.
Feedback
Unsupervised learning is accomplished through feedback and a series of rewards
and punishments. Alan Turing proposed this idea in his article on computational
intelligence [23], and it is a major theme in expert system and neural network
learning algorithms [24]. In the VT learning machine, feedback information containing the radio’s actual observations of its operation is provided. These actual
observations of the performance parameters are then compared to simulated
parameters from the WSGA. The WSGA uses the performance parameters within
its GA to hypothesize the best new radio settings. When the simulated parameters
are compared to the actual meters, the evolver in the CSM can determine how
well the simulated performance matched the actual performance. This comparison
helps the cognitive engine to understand how well its action analysis and selection
block are working. Deviations between the actual and simulated performance
measurements are penalized. The feedback and comparison mechanism allows the
evolver to update the knowledge base to better represent the environment for
improved future performance. This process requires an analysis of the model,
action, and feedback system to know what is not modeled accurately, and then it
specifically corrects this part of the knowledge base—probably by taking a new
action from another part of the knowledge base that has performed well in these
areas. This type of adaptation has been studied and used previously in GA work
through techniques such as chromosome tagging and templates or in knowledgebased genetic techniques [17]. The evolver is a highly directed adaptation mechanism to correct for specific mistakes in judgment.
Knowledge Representation
Knowledge is represented as a database of past models and actions to take, along
with past actions taken, and any estimates of the level of success associated with
these actions. The knowledge base is a set of these models and actions that allow
the WSGA to perform better as improved knowledge is gathered or generated by
the evolver. As models are observed, actions are taken that best respond to the
model’s requirements for the radio. As these correlations are learned, the cognitive
242
Cognitive Techniques: Physical and Link Layers
engine can more easily respond to situations as they arise. Likewise, past actions
are kept in short-term memory to provide the WSGA with a set of recent actions;
when a change is required, it may not have deviated too much from the immediate
past, and the recently used actions may already work to guide the WSGA. The
knowledge base is designed to provide the WSGA with as much relevant information as possible to reduce the computational complexity of the GA. If the population is primed with good solutions, the algorithm can terminate early and adapt
the radio faster.
Case-Based Decision Theory
The core of the decision-making process is rooted in CBDT [25], a technique that
grew out of economic decision-making research and is similar to the CBR used in
some AI implementations. CBR recognizes similarity between the current event
and past events in its memory, and actions taken by the machine for that past
event are used to determine the actions the machine will take in the new event
(these actions may not be the same). CBDT, in contrast, not only calculates a similarity of events in memory, but also assesses the utility of the past actions. The
added knowledge of CBDT allows the decision-maker to act autonomously, more
intelligently, and with better results than a CBR system or other memory-based
approaches. CBDT is discussed in more detail in Section 7.7.2.
The key responsibility of the CSM is to maintain the knowledge base of
model–action pairs, which ties the abstract models developed by the modeling
system with the set of actions taken by the implementation system, which is the
WSGA of Figure 7.11. The actions consist of genetic parameters such as crossover
and mutation rates; previously used chromosomes that are related to both the specific channel and network model (exploiting benefits from long-term memory) and
the most recently used chromosomes (exploiting benefits from short-term memory); and a set of objective (or fitness) functions and function weights. The actions
provided by the CSM define the fitness landscape [17] and prime the GA.
Learning
Learning is the aggregation of the above processes. This learning machine mechanism is similar to processes known to occur in human learning: sensing, acting,
reasoning, feedback, and accumulating knowledge and experience. It is a loop that
continuously improves the cognitive radio’s ability to perform. An even more
powerful technique arises when the knowledge and learning capabilities are
shared across the network. As any one radio gains knowledge about the models
243
Chapter 7
and actions to take, all radios benefit by the distribution of this knowledge.
Likewise, the computational processes involved in algorithms such as the WSGA
can be shared among many radios in the network, reducing the computations any
one radio must perform as well as increasing the speed of the algorithm. Another
benefit to distributing the cognitive engine is to share radio capabilities. For
instance, if one radio in the network has a sounder such as that described by
Rondeau et al. [26], then the modeling capabilities could be shared among all
other radios on the network without such equipment.
7.6.2 Representing Radio Parameters as Genes in a Chromosome
The previous sections of this chapter introduced the possible knobs of the PHY and
link layers that a cognitive radio can adapt, and argued that we can treat the radios
as biological systems with chromosomes that use genes to represent the different
knobs. Here, both of these thoughts are expanded to show the representation.
The greatest difficulty in the representation is the relatively large difference in
the range of possible values (settings) for each knob. With a frequency-agile radio
spanning frequencies from, say, 1 MHz to 6 GHz and a step size of 1 Hz, we must
represent approximately 6 109 discrete frequencies in the gene. In contrast, the
number of different modulation types is limited to a few dozen. Within each of
these are only a handful of modulation indices or orders. Thus, modulation offers
far fewer genetic possibilities than does operating frequency.
For the frequency representation, we could use clever programming techniques to realize different length genes for different traits; however, this quickly
becomes problematic when a 37-bit gene represents the frequency (37 bits allows
1 Hz resolution from direct current (DC) to 100 GHz)—a gene of this size is very
difficult to explore in a GA because a random search through this space is unrealistic (237 possibilities). One practical approach is to segment these into multiple
genes, where one gene provides a certain bandwidth of spectrum in which to
search and the other provides the resolution within that bandwidth. Because both
of these are genes, the search responsibility is distributed between the genes, and a
usable band and center frequency can be found faster.
Knobs with smaller search spaces can be realized with smaller genes, and some
traits can be combined. Modulation is segmented into the type of modulation—such
as amplitude shift keying (ASK), frequency shift keying (FSK), phase shift keying
(PSK), QAM, and so forth—and the order (2-, 4-, 8-, 16-point constellations, etc.).
A 16-bit gene representation (65,536 possibilities) is far too large for either type
or order, but the gene could be split into two 8-bit pieces. A system could then
244
Cognitive Techniques: Physical and Link Layers
adapt these parameters together (choosing BPSK or 16-QAM) or as two separate
genes (PSK with order 2 or QAM with order 16).
The challenge lies in scaling the GA to the full-scale factors of real-world
applications, dealing with the difference between generality and domain-specific
knowledge. Segmenting the genes into different sizes depending on traits leaves
us with the problem of deciding how to best do this. If we overshoot our domain
range, we can end up with large genes and overly complex search spaces. For
example, if we decide to have a 37-bit gene for frequency, what do we do if
we are applying the algorithm to a radio that can only realize 1–2 GHz center
frequencies? We could easily represent the frequency range with a 30-bit gene
(for just over 1 109 possibilities), but the remaining search space would be
detrimental to the algorithm. However, setting a static field of 30, or the easier-torealize 32-bit values, we could lose resolution for more complex or flexible radios.
A way around this is to use dynamic programming techniques and meta-level
intelligence. With dynamic programming, we can instruct the GA to alter its gene
size in real time for more appropriate adaptation to the current radio system.
Understanding how to adjust the gene size can come from a higher-layer intelligence that understands the radio capabilities and environmental situation to adjust
the GA behavior appropriately. Part of the responsibilities of the modeling system
and the cognitive system previously discussed in section Modeling is to realize the
meta-level intelligence. The various techniques to do this can change from platform to platform and between programming languages.
7.6.3 Multi-dimensional Analysis of the Chromosomes
Section 7.5 showed how different applications and users will have different performance criteria from the radio. This section discusses ways to capture these
effects in the GA-based cognitive engine. When evaluating each chromosome in
all of the objectives, we must find some way to analyze the performance of all
radio chromosomes in such a way as to reflect the user needs and radio performance under current channel conditions.
Again, by using dynamic programming techniques, the CSM of the cognitive
engine can learn the user and application needs and instruct the WSGA about
which objectives matter and how much they matter. This information is passed
from the CSM to the WSGA as a set of objective functions to use along with relative weighting values for each objective. For a video link, BER will receive a
smaller weight than data rate and latency, and a standard data application, such as
File Transfer Protocol (FTP) or HyperText Transfer Protocol (HTTP) will have a
245
Chapter 7
higher weighting for BER and PER. It is the job of the modeling and learning
machines to make the necessary distinctions and learn the appropriate responses.
When evaluating the chromosomes based on the relevant objectives, the most
obvious way to compare chromosomes is based on a simple summation of the
weighted objectives. The next section discusses and shows how, once again, the
trade-off between generalization and domain-specific knowledge causes us to
rethink this standard method of evaluating MODM problems.
Linear Combination of the Objective Function Results
When trying to compare different solutions, it is useful to fall back on a single
metric of evaluation. Because a MODM problem develops multiple metrics for
each objective, summing over all dimensions to present a single solution presents
obvious advantages—it is quick and simple.
Although the summation method is simple, however, it is not always wise,
especially when comparing functions of different dimensions. Cost and efficiency
are difficult to compare in a summation if efficiency has a domain that spans 0 to 1,
whereas cost can span dimensions of hundreds to millions of dollars. In these situations, a multiplicative or a ratio combination may be useful. With cost and efficiency multiplied together, we get an efficiency-weighted cost value useful in
comparing different solutions.
As the solution space increases to more than two dimensions, however, it
becomes less straightforward to combine the solutions. Summation will not work
due to the varying dimensions of each objective, and multiplication at this level of
complexity brings its own problems. Two main issues arise here. The first, and
easiest to handle, is that multiplying many dimensions of large quantity may result
in exceedingly large values that overrun the standard data structures of the computer. This problem is quite simple to avoid (section Normalizing Objectives on
normalizing the values to the value range diminishes this issue). The second, and
more complicated issue of a multiplicative combination, is that different combinations of values in each dimension can lead to the same solution.
Take, for example, a situation in which we are trying to compare two Gaussian
PDFs by trying to find the PDF that maximizes both the mean and the variance.
Which is a better solution: a mean of 5 and variance of 2, or a mean of 10 and
variance of 1? Without putting any constraints or values on the different
dimensions, both are potential solutions, and comparison is difficult because the
multiplicative combination of the objectives is 10 for both solutions. With equivalent solutions, we have a problem in deciding which one to choose; yet in some
246
Cognitive Techniques: Physical and Link Layers
cases, maximizing the mean might be more important than maximizing the variance, or the variance’s range of values might be less and so a difference of 1
between the two variances is a lot more than a difference of 5 in the mean.
With an additive combination of objectives, we can still achieve the same result
with different values of objectives. As a trivial example, consider comparing
Gaussian PDFs by which a mean of 5 and variance of 1 has the same result as a mean
of 4 and variance of 2; that is, both have linearly additive objectives of 6. In this case,
however, we can start to put some weighted value on each dimension. If maximizing
the mean were more important to the solution, we could weight the mean value by 2
and the variance value by 1, giving our example here overall solutions of 11 and 10
as opposed to 6 for both. Although we can still see problems with similar solutions,
we have flexibility in how we treat each dimension, and we have spread out the solution domain to make comparison easier. This same solution cannot be achieved with
the multiplicative approach, however, because multiplying one dimension just adds
another term into the multiplication and effectively multiplies all dimensions. The
concept of valuing objectives unequally was covered in Section 7.5.1.
Another problem with these methods is that we have been assuming we are optimizing all objectives in the same direction. That is, we are either maximizing or minimizing in each dimension. Furthermore, the range of values possible for the different
objectives may be to an extreme that a simple multiplication does not properly connote the differences in the solutions. To handle these problems, the next section presents an analysis of how normalizing the objectives can help to solve this problem.
Normalizing Objectives
We would like to try and fit our solutions into some comparable and workable
solution space. As we have seen, comparing different objectives with varying
domains is problematic. Also problematic would be trying to maximize one
dimension while minimizing another. Both of these issues are important in the
radio optimization problem. First, power consumption is a minimization problem
valued in dBm or possibly mW, and BER is a minimization in frequency of errors
per unit of received bits. The two values have large variation in their domains,
with power having values of 103, whereas BER for a data system would have
values on the order of 106 and lower. If we then add data rate, which is a maximization problem on the order of 103 (Kbps) to 106 (Mbps), the problem is
compounded. In order to compare all of these possible dimensions, we need some
methods other than the linear combination techniques discussed in section Linear
Combination of the Objective Function Results.
247
Chapter 7
For these problems, it is possible that we can normalize all the objective’s
domains to comparable ranges. For the objectives we wish to minimize, we can
take the reciprocal values and then try to maximize them. If we assume that a typical BER of 106, a power of 103 W, and a data rate of 1 Mbps are appropriate
values for our radio solution, we can divide the values by these typical values and
then invert the solutions for the power and BER objectives. We now have three
comparable values for the maximization problem.
This concept, unfortunately, works only when we have enough domain knowledge to set the typical values to normalize the solutions [10]. In communications, a
typical data link should have around 106 BER, but an audio link is sustainable up
to and around a BER of 103. In a cognitive radio, we have to account for all of
these applications and their different performance demands on the radio to properly
normalize them. This requires a lot more information than we wish to store
because the domain knowledge required is great and flexible. We need some other
way of comparing the solutions, preferably on a dimension-by-dimension basis.
Relative Tournament Evaluation
We use a relative tournament selection method similar to that proposed by Lu and
Yen [15], except that the fitness of the winner from a single comparison is scaled
by the weight associated with that objective. After all of the single comparisons in
all dimensions, the winning member, and the one that becomes a parent to the
next generation, is the one with the highest fitness. Although this method does not
guarantee that all winners are the best, or nondominated, members of the population (only better relative to their ancestors), it maintains species diversity within
the population while still pushing toward a Pareto front (see Section 7.5.3).
Diversity in the population allows different solutions to be tried and helps to
prevent the algorithm from getting stuck in a local optimum.
Unlike the linear combinations of objectives, this method compares like objectives only. There are no issues, then, of improperly comparing dimensions and
ranges. If we are trying to minimize BER, 106 is obviously better than 104
without any obfuscation by additively or multiplicatively combining these values
with data rates of 104 and 106. We are now comparing objectives of similar
dimensions and ranges. There will be no need to normalize the BER values
to compare against transmit power values of 10 dBm or 1 dBW.
If we define F(Ci) as the overall fitness of chromosome i, and wk as the weight
associated with an objective, k ∈ K, then we compare the chromosomes in each
dimension k with the algorithm shown in Eq. (7.7). The value I is 1 if the fitness
248
Cognitive Techniques: Physical and Link Layers
of a chromosome of interest is greater than another chromosome, else it is 0 (i.e.,
the chromosome is awarded points for winning the tournament).
⎧⎪1,
Ii ⎪⎨
⎪⎪0,
⎩
F (Ci ) ∑ Ii ⋅ wk
k
fi,k f j ,k
fi,k f j ,k
(7.7)
where
⎪⎧1,
I j ⎪⎨
⎪⎪0,
⎩
F (C j ) ∑ I j ⋅ wk
k
f j,k fi,k
f j,k fi,k
So the winning chromosome has its weight adjusted by the weight associated
with the objective. The weightings then help the decision-maker adjust its choices
by putting more or less emphasis on objectives as their importance to the solution
changes.
7.6.4 Relative Pooling Tournament Evaluation
In our tournament selection process, we randomly choose between two chromosomes to select one as a parent. Another method is to randomly choose two parents and a random pool of chromosomes from the remaining population [27].
Each parent will compete independently with each member of the population
pool. Whichever parent wins the most number of tournaments wins the selection
and ability to reproduce. This method leads to parents who are more fit for the
user’s needs and the communication channel within the population as a whole,
and puts a lot of selection pressure on the algorithm.
Here, we did not apply Horn et al.’s [27] method of fighting two individuals
against a subset of the population because their method, although they claim it
produces better results, calls for a larger population and more fitness comparisons.
This becomes computationally intensive and we have yet to test to see if the
improvement is worth the computational cost.
7.6.5 Example of the WSGA
This section presents the results described by Rondeau et al. [22]. The first test
conducted employed a real hardware test bed. However, these radios were Proxim
Tsunami radios, which are hardware-based platforms with limited knobs and tuning range; all adjustable PHY-layer properties are listed in Table 7.2.
Even with this limited radio platform, the GA optimization technique was
found to be viable. The scenario design consisted of a point-to-point radio link
249
Chapter 7
Table 7.2: Proxim Tsunami adaptable parameters.
Parameter
Range
Frequency
Power
Modulation
Coding
TDD
5730–5820 MHz
6–17 dBm
QPSK, QAM8, QAM16
Rate 1/2, 2/3, 3/4
29.2–91% (base station unit to subscriber unit)
QPSK: quadrature phase shift keying; TDD: time division duplex.
controlled by the WSGA and a third radio of the same type that acted as an interferer. The scenario is shown in Figure 7.12 and results are shown in Figure 7.13.
The test consisted of setting up an initial video-streaming link with high-quality
throughput (Figure 7.13(a)). When the interference started, the signal quality quickly
Figure 7.12: Initial test of WSGA hardware
platform on Proxim Tsunami radios. The test used
two radios as friendly net members and another
radio as an interference source.
Interfering Unit
Network Base
Station Unit
(a)
(b)
Network
Subscriber Unit
(c)
Figure 7.13: Results of the initial test of the WSGA hardware platform. The radio
experienced degraded performance with onset of interference and then found acceptable
waveform modifications to avoid interference.
250
Cognitive Techniques: Physical and Link Layers
dropped and became barely distinguishable (Figure 7.13(b)). The WSGA was then
run with the objectives set to minimize BER, minimize transmit power, and maximize
data rate, yet the radio was prevented from switching frequencies because that was the
easy solution and the test was for how the radio would handle its other parameters.
The result was that it automatically reestablished the video link (Figure 7.13(c)).
Although this test showed good performance in a live test, a more flexible
platform was desired, so a PHY-layer simulation of a software-defined radio
(SDR) was developed with more knobs to tune and larger tuning ranges, as listed
in Table 7.3. The resulting radio simulation’s objective functions were BER, bandwidth occupancy, spectral efficiency, power, data rate, and amount of interference.
This test had three goals: minimize spectrum occupancy, maximize throughput, and avoid interference. The weighting values for the available dimensions are
shown in Table 7.4.
Table 7.3: Simulation adaptable parameters.
Parameter
Range
Power
Frequency
Modulation
Modulation, M
PSF roll-off factor
PSF order
Symbol rate
0–30 dBm
2400–2480 MHz
M-PSK, M-QAM
2–64 M
0.01–1
5–50
1–20 Msps
Table 7.4: Simulation test conditions.
Weights
Functions
BER
Bandwidth (BW)
Spectral efficiency
Power
Data rate
Interference
Minimize spectral
occupancy
Maximize
throughput
Interference
avoidance
255
255
100
225
100
0
100
10
200
10
255
0
200
255
200
200
100
255
251
Chapter 7
The results are graphically illustrated in Figure 7.14, which shows that each
objective was met, and each test resulted in a BER of 0. The first test (Figure
7.14(a)) minimized the spectrum occupancy to 1 MHz; the second test (Figure
7.14(b)) maximized the throughput to 72 Mbps; and the third test (Figure 7.14(c))
found a solution that fit into the white space left by the interferers.
(a)
5
5
(b)
0
5
5
Magnitude (dB)
Magnitude (dB)
0
10
15
20
15
20
25
30
25
30
2400
10
35
2420
2440
2460
2480
40
2400
Frequency (MHz)
(c)
5
5
Magnitude (dB)
2440
2460
Frequency (MHz)
2480
Figure 7.14: Results from WSGA simulation
tests. The tests showed that the algorithm
could find acceptable solutions under a
variety of different interference conditions
and objective criteria.
Signal
Interferers
0
2420
10
15
20
25
30
35
40
2400
2420
2440
2460
Frequency (MHz)
2480
7.7 Advanced GA Techniques
The previous sections have introduced the GA mostly through the example of a
simple GA. This GA is only the starting point for real GAs, and adjustments have
already been made in the simple GA to make the multi-objective GA. However,
far more techniques are available to improve the performance or tailor the GA
performance to a specific problem. Rather than presenting here an entire treatise
252
Cognitive Techniques: Physical and Link Layers
on the subject of GA techniques, this section is restricted to some of the more
popular and useful techniques. The reader is referred to the literature for
details; two excellent general sources of GA techniques are Chambers [28] and
Goldberg [17].
Niching is a popular technique to ensure population diversity throughout
the GA. In a population, we wish to spread our chromosomes around the
solution space, especially if the space is multimodal (has many local optima).
A niche is a particular range within the entire solution space, and we want the
GA to maintain some presence in all niches to create many localized search
spaces in hopes of finding the global optimum. Niching has a long history in
GAs, going back to De Jong’s crowding [20] in 1975. A multiniche-crowding
algorithm is described in detail with many useful examples in Vemuri and
Cedeno [29]. Fonseca and Fleming [30], and Horn et al. [27] use niching in multiobjective GA searches. Niching is one way of maintaining diversity along the
Pareto front.
Dividing a total population into some smaller subset of populations is a technique used to reduce the computational complexity of each algorithm. These techniques have been tested to run different groups of solutions simultaneously to find
the same optimum. These methods have been done both on the same platform or
in a parallel method on multiple parallel platforms. Not only do these populations
allow more searches to occur, they also have migrating populations between the
“islands” to share highly fit members and maintain diversity. Many researchers
have studied this issue [17, 31, 32].
GAs are sensitive to parameters such as crossover and mutation rate, population size, and replacement size, but nothing stops a GA from altering its parameters throughout its lifetime [33, 34]. One method is to monitor the performance
improvement over the generations. As the improvement slows down, we might
have converged to a local optimum or a global optimum. Not being sure, it would
be wise to keep looking, for a while at least, but an increased mutation rate would
enhance the exploration features of a GA to let it search in new areas of the search
space. If the GA was caught in a local optimum, a high mutation rate will help
break out of that cycle to find the path toward a new optimum, hopefully the
global optimum this time.
7.7.1 Population Initialization
Although these different methods have been proposed to enhance the performance of GAs, overall only a rather small subset focuses on the population’s
253
Chapter 7
initialization strategy. A few good representatives of the initialization methods
include biasing the initial population using domain knowledge [35], creating unbiased or diverse populations [36], and using case-based initialization techniques
[37]. The case-based systems, in particular, have been shown to lend themselves
to successful operation in changing situations, or “anytime learning,” as Ramsey
and Grefenstett [37] define it. These systems are very similar to the online learning we wish to accomplish. The technique in this chapter is similar in that it presents an advanced understanding of the operation and benefits that case-based, or
memory-feedback, GAs offer.
7.7.2 Priming the GA with Previously Observed Solutions
The method of priming the GA with previously observed solutions falls under the
category of case-based intelligence in the AI literature. Here, we can look into the
two methods discussed in section Case-Based Decision Theory: CBR [38] and
CBDT [25]. The application of CBDT to the initialization problem, as presented
here, adds to other case-based methods [37]. Those systems are usually inspired
by the field of CBR [38], which uses case look-up based on similarity, but the
application in this chapter is a decision-making theory based in both similarity
and utility [22]. The added dimension aids the decision-making process such that
cases in memory are not only similar to the current situation, but they have also
performed very well in past situations. This section presents the basics behind the
CBDT application to GAs and its improvement to online GA optimization systems. In doing so, it provides a more complete analysis of how the memory system works.
CBDT was introduced by Itzhak Gilboa and David Schmeidler in the mid1990s [25]. The theory is a method of accessing past knowledge to help make
decisions for present situations. The theory defines the set of cases to include all
possible problems, actions, and results in which a decision-maker contains a
memory M, which is a subset of the total number of possible cases, C. Current
decisions are based on the knowledge of past cases by defining the amount of similarity of the current situation to the past cases and the utility of each case in memory. The case that maximizes the desirability based on the utility and similarity is
chosen as the case that best applies to the current situation.
Formally, CBDT is defined to have q ∈ P set of problems, a ∈ A set of actions,
and r ∈ R set of outcomes. A case, c, is a tuple of a problem, an action, and a
result such that c ∈ C, where C ∈ P A R. Furthermore, memory, M, is formally defined as a set of cases c currently known such that M C.
254
Cognitive Techniques: Physical and Link Layers
Similarity is defined in Eq. (7.8) as how similar two decision problems are to
each other:
s : P P → [0,1]
(7.8)
Utility is defined in Eq. (7.9) as the level of desirability the outcome represents:
u:R → ℜ
(7.9)
We concentrate on the simplest formulation, where the most desirable act is chosen from the set of known acts in memory by using the product of similarity and
utility [25], as defined in Eq. (7.10):
U (a) s( p, q)u(r )
where (q, a, r ) ∈ M
(7.10)
For a given problem, p, this equation determines the desirability of action a. The
action that maximizes this function is the chosen action to respond to problem p.
Section 7.7.3 introduces how the GA is aided by CBDT.
7.7.3 CBDT Initialization of GAs
The Theory
If the cognitive radio system builds a memory database that represents the outcomes of continually running the GA for real-time optimization, then as the system learns and optimizes, each new optimization would take significantly less
time to run, or, conversely, would find a more optimal solution in the same
amount of time. For the GA, we represent a new input as the problem p. We have
previously observed problems, q, in a database associated with a set of chromosomes, a, used to model the problem along with their fitness values, r.
When the system receives this new problem, it finds the action within its
memory that maximizes the product of the utility and similarity functions. The
action is a set of chromosomes that partially initializes the GA population while
the rest of the population is randomly generated.
Chou and Chen [36] discuss the use of uniformly generating a population to
have an equal representation of all points in space, which should lead to at least
one chromosome close to the global optimum. As such, by initializing with both
the case-based solutions and randomly generated solutions, we maintain some
diversity for proper searching while focusing the search in a region of presumably
high fitness, the concept illustrated in Figure 7.15.
255
Chapter 7
Figure 7.15: Search space. The square is the
new target optimum, the triangle is the
previous target optimum, and the stars are
the initialized population. The initialized
solutions from the CBDT system (dark gray)
concentrate around the previous optimum
while other initial solutions (white) are spread
out randomly to cover the whole search space.
The initial population is of size Np, where Ma members are initialized from the
case-base and Np Ma members are randomly initialized. In order to maintain the
diversified population, we ensure that Ma is always less than Np.
We define the action associated with each case in memory to consist of a number of chromosomes, M, and the results are therefore the fitness of each chromosome as fed back from the GA along with the respective chromosome. The utility
of the case k, shown in Eq. (7.11), is then defined as one minus the sum of all of
the solutions’ fitnesses stored in each case:
M
u(rk ) 1 ∑ fki
(7.11)
i1
The winning case is the case that maximizes the product of the similarity and
utility functions as defined by Eq. (7.10).
From here, we define two initialization methods. First, some number of solutions, M 1, are associated with each case, and all M solutions are used to initialize the GA. The other method sets M 1 and the solution of the winning case is
used to initialize the GA, along with some number of surrounding cases. Other
ways to mix these two techniques exist, but are not discussed here.
The Implementation
The design of the information flow is shown in Figure 7.16. Here, the input is
compared to the cases in memory through both similarity and utility calculations.
The chromosomes (action space) of the winning case are sent to the GA as well as
the input. The GA then runs to optimize the result within a set period of time. The
GA outputs a solution and its fitness value, which are both fed back to memory.
The initialization can happen in two ways. Either the case-base stores multiple
solutions to one case in memory and sends all these solutions to the GA, or the
256
Cognitive Techniques: Physical and Link Layers
Figure 7.16: System diagram of CBDT initialization of GA.
case-base stores one solution per case and sends solutions from multiple cases to
the GA.
The number of chromosomes associated with any case in memory is limited to
some number, M, and when a new chromosome is inserted because it is locally
optimal for current conditions, the oldest chromosome (or least suitable) is
replaced. When inserting a new case to the finite memory space of size N, the oldest case in the memory is forgotten and replaced by the most recently observed
case. The new memory case is then associated with both its most recent solution
and the solutions of the case originally selected from memory if M 1. Through
this update procedure, the memory is kept current and positive actions used in the
past are preserved.
Other possible methods for replacing cases and solutions include replacing the
case with the worst overall utility or a case in memory that has been used less frequently than the others. Both long-term and short-term memories with different
forgetting methods to exploit both time (short-term) and utility (long-term) benefits have been considered.
Two important issues to consider are the size of the memory and the initialization method. The size of the memory has a direct effect on the computational
complexity of the decision-maker. The choice of the memory size should reflect
the trade-off between a large memory that will add to the time required to both
learn and process the input signal versus the increased efficiency to find solutions
similar to current conditions. If this added time is too great, it begins to break
down the benefits achieved through the system. The initialization method is the
choice between associating multiple solutions with a single case and sending only
257
Chapter 7
those solutions to the GA, or associating only a single solution per case and sending multiple cases to the GA.
7.8 Need for a Higher-Layer Intelligence
7.8.1 Adjusting Parameters Autonomously to Achieve Goals
An important goal of the cognitive engine’s higher-layer intelligence is to
autonomously set the weights for the GA in order to both reduce limitations in
design and remove operator interaction with the radio. The preference is that the
cognitive engine has the ability to set weights on its own. By observing the user
and the needs of the system, the weights should be able to adjust themselves
accordingly, with purpose and intelligence. The system designer could establish a
preset table of weights for different applications, but these presets are limited by
the imagination of the designer and current understanding of the uses of the cognitive radio. If new situations and needs arise in the field, creative solutions should
be developed as they are needed.
The cognitive engine should reflect the user’s needs without burdening the
user with developing the weights themselves or constant communication between
the radio and the user. Purpose, intent, and lower-layer needs should be inferred
from the higher layers and not from the user answering questions from the radio
such as, “Would you like to increase your data rate?”
The last requirement, to infer user and application needs, is possibly the most
challenging concept introduced so far and is a rich subject for research. The current design uses a simplified model, as discussed in Section 7.8.2.
7.8.2 Rewards for Good Behavior and Punishments for Poor Performance
In his landmark article, “Computing Machinery and Intelligence,” Alan Turing
[23] recalls his British public school days, where discipline was taught through
physical punishment, which he himself both gave and received, but we can adapt
this and think of it as positive and negative reinforcement. Provided the input from
the user and the radio, and having produced the WSGA output, the engine now
requires a means to tie the information together and learn from its past behavior,
both its mistakes and successes.
A reinforcement loop must exist between the WSGA solution based on its
simulated meters when solving for the objectives and the actual meters observed
by the radio while operating under the parameters set by the WSGA. If the
WSGA’s optimization was incorrect, the system should be adjusted (punished) for
258
Cognitive Techniques: Physical and Link Layers
improved performance next time. If the WSGA’s optimization was correct, the
system should be rewarded to make sure it does the same in future situations. Two
ways of distributing these rewards and punishments are discussed next.
Rewards and Punishments Can Be Radio Algorithm-Inflicted
The cognitive radio must have some sense of its performance and adjust itself
accordingly to improve its performance. The cognitive radio must know its current internal performance and capabilities in its approach to intelligent optimization, an approach that has been called “self-aware” in current cognitive radio
circles. This does not mean that the radio is self-aware in any sense that humans
are self-aware; instead, it refers to the radio’s ability to keep track of its performance and available resources (battery life, geography, computational power, usecase scenario) and alter its behavior to account for needs and changes required by
this understanding. On this level, the radio will use the knowledge of itself in the
optimization problem, where battery life and computational power of waveform
generation can play major roles in the radio performance.
On another level, the radio must be conscious of its external performance
measures. When the GA adjusts the radio, the adjustments are expected to produce some known result, such as a target BER or data rate. The cognitive radio
then monitors the actual observed performance and will either reward or punish
itself for differences between the actual and simulated performances.
For this to work, the radio must be able to read the same meters used in the
GA optimization process. The farther a meter deviates from the simulated performance, the more the cognitive radio process is punished. Again, punishment
results in changes to the radio decision-making process to help it change itself for
improved performance later.
Because the radio will be able to measure some meters itself, such as SINR
and PER, the amount of deviation between these observed meters and the
WSGA’s simulated meters should tell the cognitive engine how best to adapt. It
would do this by asking questions such as, “Is this PER right for this SINR?” If
the answer is “yes” then the PER can possibly be improved by increasing transmitter power. If the answer is “no” (the PER is too high for the given SINR), then
the problem is interference or multipath, and the PER can possibly be improved
by changing frequency or modulation parameters. This method suggests the use
of an expert system AI technique [24]. By programming just a few basic rules of
radio operation, we can infer a lot and adjust the system appropriately. Expert systems are difficult to apply, especially to a general problem setting, because they
259
Chapter 7
use very specific, preprogrammed knowledge. Mixing expert systems with fuzzy
logic [39, 40] allows much more flexible reasoning. Also, an implementation of
an expert system in this case should be very general, using only the most basic
wireless communication rules to make inferences. Other techniques, such as evolutionary strategies [24], could then be tiered with the expert system to make more
creative assertions about the problem and how to fix it.
Rewards Can Be User-Inflicted
Although we advocate minimizing interaction between the user and the radio,
this can be helpful in teaching the cognitive radio. The self-reflexive method just
described is unsupervised learning, but sometimes a supervised learning process is
necessary, especially when trying to tailor behavior to a particular user. Using the
same punishment metric of credit deduction as used above, credits will be altered
based on user satisfaction. This is another area of great interest and research in the
cognitive radio community. Rather than promoting a particular method here, Section
7.9 looks to the field human–computer interface (HCI) for how to best facilitate the
interaction and collect the data required to perform the credit adjustments.
7.9 How the Intelligent Computer Operates
This chapter has discussed many aspects we expect to see from a fully functional
cognitive radio capable of altering the PHY and MAC layers with respect to the
user and external environment. It has discussed the cognitive decision-making
process using case-based theory and GAs to solve the multi-objective optimization problem. It has also referred to concepts of user and channel modeling, which
require some advanced computer techniques, and mentioned the need for selfawareness in the radio. In the end, though, there is no single method of AI capable
of performing all possible tasks required by the cognitive radio. Therefore, we
look to a tiered approach of computational intelligence and machine learning
algorithms to perform their tasks to create a fully cognitive radio. This section
revisits these concepts to show how the collection of techniques creates a cognitive engine capable of controlling and adapting a cognitive radio.
Ultimately, all cognitive radio work really comes down to the multi-objective
optimization problem, which is that the MODM search space:
1. Has no utopian point; that is, it has no point that fully optimizes all objectives.
2. Is nonlinear and nonconvex; instead, it is a very complex plane to model with
complex relationships between all inputs and outputs.
260
Cognitive Techniques: Physical and Link Layers
All other aspects of the cognitive radio are inputs to the MODM problem in order
to either set up the optimization problem or speed up convergence. (Refer back to
Figure 7.1 for the graphical representation of the items discussed here.)
7.9.1 Sensing and Environmental Awareness
The radio must first understand what is required of it and what it must be capable
of, thereby defining the optimization space. This requires sensing or awareness on
four levels: (1) recognizing the needs of the user, (2) understanding the limitations
imposed on the radio operation by the channel and external environment, (3) realizing its own limitations in flexibility and power, and (4) conforming to local regulations and policy. Although this chapter discusses a few implementation methods,
this is not the central focus, and they are only briefly covered here.
User Awareness
The user domain provides subjective information to the cognitive radio, by which
the cognitive radio is given instructions on how to behave and what to optimize.
User awareness could be either passive or active. Passive awareness requires
learning user needs and behavior by inferring concepts from the applications
being used and data being sent through the radio. Active learning is done through
direct communication with the user and HCI techniques to collect information.
Inference can be accomplished by sampling the packet headers as well as the
packet and bit rate required by the application sending the data. Such learning
could be accomplished using case-based techniques or even online, unsupervised
neural network learning through bidirectional associative memories [41]. These
techniques could learn that a header containing the protocol values for File Transfer
Protocol (FTP), Transmission Control Protocol (TCP), Internet Protocol (IP), and so
forth is a file transfer application requiring full packet fidelity, retransmission for
failed packets, and high data rate with acceptable tolerances in latency. A similar
packet that contains a protocol value for Research ReSerVation Protocol (RSVP)
instead of FTP would be remembered as dealing with real-time operation, which
would require different services. Some of these protocols and applications can be
preprogrammed into the user modeling system, but a true unsupervised learning
machine for new and unknown protocols required by the user is preferred.
Environmental Awareness
The external, RF, and geographical domains provide the environmental context.
Sensing and modeling the channel can be done in a number of ways. Sending
261
Chapter 7
packets of training data through a channel and observing the bit errors and burst
error rates is one method. Another uses channel sounding to capture a snapshot of
the channel’s propagation information [26]. Other methods exist for calculating
noise floor and interference power to understand and recognize the presence of
other users and radios [42].
Neural networks are well known for their ability to classify information and
for pattern recognition. Neural networks have been used to classify modulation
types in signals with great accuracy. Understanding not only how many other
radios are nearby, but also the types of communication the radios employ can be
of great use in determining the optimization. If a frequency is in use by another
radio network, finding an orthogonal modulation, antenna polarization, and so on
could lead to highly efficient spectrum sharing [43–45].
External radio and propagation information is also useful when looking at
the current internal radio resources. If the radio is operating in a clean channel
with few interferers, then a high-order, narrowband modulation technique
could be chosen over advanced spread spectrum signals to reduce computational
complexity (and therefore prolong the battery life) while providing high data
rates.
7.9.2 Decision-Making and Optimization
Using the knowledge gathered in the sensing mechanisms, the cognitive radio
must then make a decision as to how to optimize radio performance. The optimization problem is multi-objective and solved by using the well-known GA.
GAs are not, of course, the only method to do this, but they are well understood
and easy to implement as well as flexible and robust.
7.9.3 Case-Based Learning
We now have a modeling and adaptation system, but we now need some way to
tie them together. A knowledge-based system to store information about the modeled environment can be tied to a set of objectives to be solved by the GA. Another
entry to the knowledge base is memory of past actions taken by the GA. When
similar environmental situations are observed by the modeling systems, similar
actions should produce comparable results. If these past solutions gave positive
results, we can aid the GA optimization process by starting with this knowledge.
GAs are notorious for their length of time to find an optimal solution, but this
chapter has shown how this knowledge-based approach reduces the problem quite
nicely to a localized optimization search.
262
Cognitive Techniques: Physical and Link Layers
7.9.4 Weight Values and Objective Functions
The cognitive radio bases its decisions on a set of metrics to best represent the
requirements of the whole radio system, the network, and the user. To do this
properly, the decision-maker must value objectives differently. In the optimization
process, different needs of the radio should be handled by using different objectives. The objectives used for a given optimization problem can be learned
through successes and failures, and the weights associated with each objective
can likewise be altered to best represent the situation.
7.9.5 Distributed Learning
Learning situations arise through modeling of the environment, inferring needs of
the user, or simply asking the user for input. However, another way of learning
that could increase the power of the cognitive radio greatly is peer learning.
Children learn through their own play [46] and through teacher instruction; they
also learn from each other [47]. Similarly, what one radio has already experienced
and learned, it can share with newer, inexperienced radios, which removes their
need for trial and error, possibly successive mistakes, and greatly accelerates optimal network performance behaviors.
7.10 Summary
The goal of a cognitive radio is to optimize its own performance and support its
user’s needs. A radio is optimized when it achieves a level of performance that
satisfies its user’s needs while minimizing its consumption of resources such as
occupied bandwidth and battery power. The intelligent core of a cognitive radio
exists in the cognitive engine, which performs the modeling, learning, and optimization processes necessary to reconfigure the communication system in which
the radio operates.
The first problem in dealing with cognition in a system is to understand (1)
what information the intelligent core must have and (2) how it can adapt. In radio,
we can think of the classical transmitters and receivers as having adjustable control parameters (knobs) that control the radio’s operating parameters, and observable metrics (meters) that measure its performance. The knobs are any of the
parameters that affect link performance and radio operation. Meters are indicators
of performance on a particular level; thus, at the link layer, PER is an important
metric.
The basic process followed by a cognitive radio is this: it adjusts its knobs to
achieve some desired (optimum) combination of meter readings. Rather than
263
Chapter 7
randomly trying all possible combinations of knob settings and observing what
happens, it makes intelligent decisions about which settings to try and observes
the results of these trials. Based on what it has learned from experience and on its
own internal models of channel behavior, it analyzes possible knob settings, predicts some optimum combination for trial, conducts the trial, observes the results,
and compares the observed results with its predictions as summarized in an adaptation loop.
Without a single-objective function measurement, we cannot look to classic
optimization theory for a method to adapt the radio knobs. Instead, we can analyze the performance using MODM criteria. MODM theory allows us to optimize
in as many dimensions as we have objective functions to model. Cognitive radio
operation requires an MODM algorithm capable of robust, flexible, and online
adaptation and analysis of the radio behavior. The clearest method of realizing all
of these needs is the GA, which is widely considered the best approach to MODM
problem-solving. In it, we represent the radio parameters as genes in a chromosome and select the fittest chromosomes through a process called relative tournament evaluation.
The primary goal of the cognitive engine is to optimize the radio, and the secondary functions are to observe and learn in order to provide the knowledge
required to perform the adaptation. A cognitive radio becomes a learning machine
through a tiered algorithm structure based on modeling, action, feedback, and
knowledge representation, as shown in the cognition loop of Figure 7.10. In the
Virginia Tech cognitive engine, these functions are realized through the algorithmic structure of Figure 7.11. Its main parts are the CSM, responsible for learning,
and the WSGA, which handles the behavioral adaptation of the radio based on
what it is told to do by the CSM. The modeling system observes the environment
from many different angles to develop a complete picture. The CSM holds two
main learning blocks: the evolver and the decision-maker, which takes feedback
from the radio that allows the evolver to properly update the knowledge base to
respond to and direct system behavior.
The algorithm structure discussed in this chapter provides the groundwork for
solving the problems of the PHY- and MAC-layer adaptation. Many areas are left
to explore in the machine learning for these layers, and many potential improvements and research directions have been outlined and discussed.
Bringing all of these ideas together is one of the great challenges of cognitive
radio. By its nature, it is a multidisciplinary problem, just as the subjects of the
other chapters in this book suggest. We need solutions that can mix these disciplines into a coherent system solution that learns and adapts to the user to satisfy
264
Cognitive Techniques: Physical and Link Layers
all of the problems encountered when attempting to optimize the point-to-point
link, as the PHY- and MAC-layer algorithms presented here do, as well as the endto-end solution of a cognitive radio network—all the while both enhancing and
protecting the regulatory structure of spectrum allocation. This chapter has provided an analysis of part of this solution and presented one possible approach to it.
Acknowledgments
The authors of this chapter wish to thank all of the researchers, colleagues, and
friends that have contributed to our work. Specifically, we are pleased to recognize the members of the Virginia Tech Research Group, including Ph.D. students
Bin Le, David Maldonado, and Adam Ferguson; master’s students David
Scaperoth and Akilah Hugine; and faculty members Allen MacKenzie and
Michael Hsiao. Finally, a very big thank you goes to three former colleagues who
helped start this research: Christian Rieser, Tim Gallagher, and Walling Cyre.
References
[1] T.W. Hazlett, “Spectrum Tragedies,” Yale Journal on Regulation, Vol. 22, 2005,
pp. 2–5.
[2] C.E. Shannon, “Programming a Computer for Playing Chess,” Philosophical
Magazine, Vol. 41, 1950, pp. 256–275.
[3] “Open Systems Interconnection—Reference Model (OSI-RM). FED-STD-1037C:
Telecommunications: Glossary of Telecommunication Terms,” National
Communications Systems (NCS), 1996.
[4] J. Hawkins, On Intelligence, Times Books, New York, 2004.
[5] D.L. Decker, “Computer Evaluation of the Complementary Error Function,”
American Journal of Physics, Vol. 43, 1975, pp. 833–834.
[6] M. Abramowitz and I.A. Stegun, Handbook of Mathematical Functions with
Formulas, Graphs, and Mathematical Tables, Dover, New York, 1972.
[7] J.G. Proakis, Digital Communications, 4th Edition, McGraw-Hill, New York, 2000.
[8] S. Zionts, “Multiple Criteria Mathematical Programming: An Overview and Several
Approaches,” in Mathematics of Multi-objective Optimization, P. Serafini (Ed.),
Springer-Verlag, Wien, 1985, pp. 227–273.
[9] J.D. Schaffer, “Multiple Objective Optimization with Vector Evaluated Genetic
Algorithms,” Proceedings of the International Conference on Genetic Algorithms,
Vol. 1, 1985, pp. 93–100.
[10] E. Zitzler and L. Thiele, “Multiobjective Evolutionary Algorithms—A Comparative
Case Study and the Strength Pareto Approach,” IEEE Transactions on Evolutionary
Computation, Vol. 3, 1999, pp. 257–271.
265
Chapter 7
[11] P. Fleming, “Designing Control Systems with Multiple Objectives,” IEE
Colloquium Advances in Control Technology, 1999, pp. 4/1–4/4.
[12] Y. Lim, P. Floquet and X. Joulia, “Multiobjective Optimization Considering
Economics and Environmental Impact,” ECCE2, Montpellier, 1999.
[13] P. Pedregal, Introduction to Optimization, Springer, New York, 2004.
[14] C.M. Fonseca and P.J. Fleming, “Genetic Algorithms for Multiobjective
Optimization: Formulation, Discussion, and Generalization,” Proceedings of the
International Conference on Genetic Algorithms, Vol. 5, 1993, pp. 416–423.
[15] H. Lu and G.G. Yen, “Multiobjective Optimization Design via Genetic Algorithm,”
IEEE Proceedings of the International Conference on Control Applications, 2001,
pp. 1190–1195.
[16] T. Hiroyasu, M. Miki and S. Watanabe, “Distributed Genetic Algorithms with a
New Sharing Approach in Multiobjective Optimization Problems,” IEEE Proceedings of the Congress on Evolutionary Computation, Vol. 1, 1999, pp. 69–76.
[17] D.E. Goldberg, Genetic Algorithms in Search, Optimization, and Machine
Learning, Addison-Wesley, Reading, MA, 1989.
[18] R. Spillman, “Solving Large Knapsack Problems with a Genetic Algorithm,” IEEE
Proceedings of the Systems, Man and Cybernetics, Vol. 1, 1995, pp. 632–637.
[19] M.R. Garey and D.S. Johnson, Computers and Intractability: A Guide to the Theory
of NP-Completeness, W.H. Freeman & Company, New York, 1979.
[20] K.A. De Jong, An Analysis of the Behavior of a Class of Genetic Adaptive Systems,
Ph.D. Dissertation, University of Michigan, Ann Arbor, MI, 1975.
[21] http://www.darpa.mil/ato/programs/xg
[22] T.W. Rondeau, B. Le, C.J. Rieser and C.W. Bostian, “Cognitive Radios with
Genetic Algorithms: Intelligent Control of Software Defined Radios,” Software
Defined Radio Forum Technical Conference, Phoenix, AZ, 2004, pp. C-3–C-8.
[23] A.M. Turing, “Computing Machinery and Intelligence,” Mind, Vol. 59, 1950,
pp. 433–460.
[24] M. Negnavitsky, Artificial Intelligence: A Guide to Intelligent Systems, AddisonWesley, Harlow, England, 2002.
[25] I. Gilboa and D. Schmeidler, A Theory of Case-Based Decisions, Cambridge
University Press, Cambridge, 2001.
[26] T.W. Rondeau, C.J. Rieser, T.M. Gallagher and C.W. Bostian, “Online Modeling of
Wireless Channels with Hidden Markov Models and Channel Impulse Responses
for Cognitive Radios,” International Microwave Symposium, Fort Worth, TX,
2004.
[27] J. Horn, N. Nafpliotis and D.E. Goldberg, “A Niched Pareto Genetic Algorithm for
Multiobjective Optimization,” IEEE Proceedings of the World Congress on
Computational Intelligence, Vol. 1, 1994, pp. 82–87.
[28] L. Chambers, Practical Handbook of Genetic Algorithms: New Frontiers, CRC
Press, Boca, Raton, FL, 1995.
266
Cognitive Techniques: Physical and Link Layers
[29] V.R. Vemuri and W. Cedeno, “Multi-niche Crowding for Multi-modal Search,” in
Practical Handbook of Genetic Algorithms: New Frontiers, Vol. 2, L. Chambers
(Ed.), CRC Press, Boca, Raton, FL, 1995, pp. 5–29.
[30] C.M. Fonseca and P.J. Fleming, “Multiobjective Optimization and Multiple
Constraint Handling with Evolutionary Algorithms. I. A Unified Formulation,”
IEEE Transactions on Systems, Man and Cybernetics, Vol. 28, 1998, pp. 26–37.
[31] J.P. Cohoon, W.N. Martin and D.S. Richards, “Punctuated Equilibria: A Parallel
Genetic Algorithm,” Proceedings of the Second International Conference on
Genetic Algorithms, Vol. 1, 1987, pp. 148–154.
[32] W.-Y. Lin, T.-P. Hong and S.-M. Liu, “On Adapting Migration Parameters for
Multi-population Genetic Algorithms,” IEEE Proceedings of the Systems, Man and
Cybernetics, Vol. 6, 2004, pp. 5731–5735.
[33] M. Srinivas and L.M. Patnaik, “Adaptive Probabilities of Crossover and Mutation in
Genetic Algorithms,” IEEE Transactions on Systems, Man and Cybernetics, Vol. 24,
1994, pp. 656–666.
[34] Y.J. Cao and Q.H. Wu, “Convergence Analysis of Adaptive Genetic Algorithms,”
IEE Proceedings of the Genetic Algorithms in Engineering Systems: Innovations
and Applications, Vol. 2, 1997, pp. 85–89.
[35] J. Arabas and S. Kozdrowski, “Population Initialization in the Context of a Biased
Problem—Specific Mutation,” IEEE Proceedings of the Evolutionary Computation
World Congress on Computational Intelligence, 1998, pp. 769–774.
[36] C. Chou and J. Chen, “Genetic Algorithms: Initialization Schemes and Genes
Extraction,” IEEE Proceedings of the Fuzzy Systems, 2000, pp. 965–968.
[37] C.L. Ramsey and J.J. Grefenstette, “Case-Based Initialization of Genetic
Algorithms,” Proceedings of the Fifth International Conference on Genetic
Algorithms, Vol. 5, 1993, pp. 84–91.
[38] J. Kolodner, Case-Based Reasoning, Morgan Kaufmann Publishers, San Mateo,
CA, 1993.
[39] M. Black, “Vagueness: An Exercise in Logical Analysis,” Philosophy of Science,
Vol. 4, 1937, pp. 427–455.
[40] L.A. Zadeh, “Fuzzy Sets,” Information and Control, Vol. 8, 1965, pp. 338–353.
[41] B. Kosko, “Bidirectional Associative Memories,” IEEE Transactions on Systems,
Man and Cybernetics, Vol. 18, 1988, pp. 49–60.
[42] T. Mo and C.W. Bostian, “A Throughput Optimization and Transmitter Power
Saving Algorithm for IEEE 802.11b Links,” IEEE Proceedings of the WCNC,
Vol. 1, 2005, pp. 57–62.
[43] B. Le, “Modulation Identification Using Neural Network for Cognitive Radios,”
Software Defined Radio Forum Technical Conference, Anaheim, CA, 2005.
[44] A.K. Nandi and E.E. Azzouz, “Algorithms for Automatic Modulation Recognition
of Communication Signals,” IEEE Transactions on Communications, Vol. 46, 1998,
pp. 431–436.
267
Chapter 7
[45] V. Cheung, K. Cannons, W. Kinsner and J. Pear, “Signal Classification Through
Multifractal Analysis and Complex Domain Neural Networks,” IEEE Proceedings
of the CCECE, Vol. 3, 2003, pp. 2067–2070.
[46] J. Piaget, Biology and Knowledge, University of Chicago Press, Chicago, 1971.
[47] L.S. Vygotsky, Mind in Society: The Development of Higher Psychological
Processes, Harvard University Press, Cambridge, MA (original works published
1930, 1933, 1935), 1978.
268
CHAPTER 8
Cognitive Techniques: Position Awareness
John Polson
Bell Helicopter, Textron Inc., Fort Worth, T X
Bruce A. Fette
Communications Networks Division, General Dynamics,
C4 Systems, Scottdale, AZ, USA
8.1 Introduction
For cognitive radio (CR) to reach its full potential as an efficient member of a network or as an aid in users’ daily tasks, and even to conserve the precious spectrum
resource, a radio must primarily know its position and what time it is. From position
and time, a radio can: (1) calculate the antenna pointing angle that best connects to
another member of the network; (2) place a transmit packet on the air so that it
arrives at the receiver of another network member at precisely the proper time slot to
minimize interference with other users; or (3) guide its user in his or her daily tasks
to help achieve the user’s objectives, whether it be to get travel directions, accomplish tasks on schedule, or any of a myriad of other purposes. Position and time are
essential elements to a smart radio. Furthermore, from position and time, velocity
and acceleration can be inferred, giving the radio some idea about its environment.
Geolocation applications are also a key enabling technology for such applications as spatially variant advertisement, spatially aware routing, boundary-aware
policy deployment, and space- and time-aware scheduling of tasks. These capabilities enable a CR to assist its user to conveniently acquire goods and services as
well as to communicate with other systems using minimal energy (short hops) and
low latency (efficient directional propagation of packets through a network). Geolocation applications in a CR enable the radio to be carried throughout the world
and used without any manual adjustment or modification to maintain compliance
269
Chapter 8
with local regulations. Finally, space- and time-aware scheduling of tasks improves
the efficiency of CR operations by managing vital resources and accomplishing
goals “at the right place” and “just in time.”
Section 8.2 covers various techniques for a CR to geolocate itself and other
systems, and Section 8.3 addresses network-aided position awareness. Section 8.4
presents the mathematics of time of arrival (ToA), time difference of arrival (TDoA),
and frequency difference of arrival types of systems. Section 8.5 discusses the
method of converting global positioning system (GPS) x-, y-, z-coordinate locations into latitude and longitude, or geopolitical region localization. Section 8.6
provides an example of boundary decisions, and an example for 911 geolocation
for first responders is given in Section 8.7. Section 8.8 discusses interfaces with
other cognitive subsystems. Finally, Section 8.9 summarizes this chapter.
8.2 Radio Geolocation and Time Services
Several radio transmission services are located throughout the world to aid in
geolocation and accurate time tracking. The best-known system is GPS1 [1],
based on a constellation of satellites constantly broadcasting time and position
(see the next section for details on GPS). The National Institute of Standards and
Technology (NIST) radio station, call sign WWV, which continuously broadcasts
time with high accuracy, is somewhat well known in the Western Hemisphere.
(WWV consists of stations call signs WWVB and WWVH [2–4]). However, without knowing position, it is difficult to determine how long the transmission took to
propagate to the receiver, and thus it renders only a coarse time.
Less well known are the very high frequency (VHF) omnidirectional ranging
(VOR) transmitters used by aircraft to locate current position [5], LOng RAnge
Navigation (LORAN) used by ships at sea to calculate position, and the geopositional services of cellular telephone systems [6]. Also quite widely deployed is
geolocation by wireless local area network (WLAN) Internet Protocol (IP)
address [6], in which the IP address of a WiFi® access point is directly translated
into a geolocation [7]. The most recently published technique is the use of television (TV) broadcasts for time, frequency, and position [8]. The attempt to address
these location techniques in this section will be presented only to the degree that it
is clear that a software-defined radio (SDR) can participate in these time and location processes, and can thereby know geolocation and time sufficiently to aid its
network, to aid its use of spectrum, and to aid its user.
1
Similar systems have also been launched by Europe (Galileo) and by Russia (Glonass).
270
Cognitive Techniques: Position Awareness
8.2.1 GPS
GPS is without a doubt the best-known location system in the world. GPS is a
satellite navigation system funded and controlled by the US Department of
Defense (DoD) and the Department of Commerce (DoC). The system comprises a
constellation of satellites, ground control stations, and GPS receivers. At most
points on Earth (other than in the deep urban canyons between skyscrapers), there
is a high probability of line-of-sight (LOS) contact with multiple GPS satellites.
Given LOS with four or more satellites, three-dimensional (3-D) position and
time can be measured.
Satellite System Architecture
The GPS system is readily divided into three segments: space, control, and user.
Space Segment
Not counting orbiting spares, there are normally 24 active GPS satellites. The
orbital period is nominally 12 hours. The 24 satellites are distributed evenly in six
orbital planes with 60-degree separation between each of the four satellites in
each plane. The inclination is about 55 degrees off the equator. This geometric
distribution provides between five and eight satellites in view from any point on
Earth. A line-of-sight (LOS—meaning unobstructed) view of four or more satellites is needed to process the signals and calculate location.
Control Segment
Ground tracking stations are positioned worldwide to monitor and operate the constellation of GPS satellites. The master control station is located at Schriever Air
Force Base in Colorado. The stations monitor the satellites’ signals, incorporate
them into orbital models, and calculate ephemeris data that are transmitted back to
the space vehicles. Ephemeris data are, in turn, transmitted to GPS receivers.
User Segment
GPS receivers and their operators form the user segment. The receivers process
the signals from four or more satellites into 3-D position and time. In the differential mode, a reference GPS receiver communicating with another GPS receiver,
where the position of one node is known to high accuracy, can then improve upon
the inherent accuracy of another stand-alone GPS receiver. GPS receivers also
produce a precise one-pulse-per-second signal.
271
Chapter 8
The one-pulse-per-second timing pulse as seen by the receiver experiences a
propagation time delay from each satellite. Because each satellite is at a different
distance from the satellite, these time pulses are seen at different times for different
satellites. Once the receiver solves for its location and the corresponding propagation time delay, it can then attribute an additional delay to all of these pulses that
will exactly align all of them to when the next pulse should occur. Thus, a sophisticated receiver can reproduce one-pulse-per-second time-aligned pulses on all
ground units over a large regional area by adjusting the pulses to compensate for
propagation delay, and then identify the time that corresponds to those pulses, just
as if the atomic clock in the satellite were connected by a wire to the receiver. As a
result, all receiver units will be able to have precision time, regardless of their position. Given the internal logic that performs this alignment, the precision of the output pulse is reduced to around 340 nanoseconds (ns, equal to 109 seconds).
The operators of GPS receivers include military users as well as civilian users.
The military operators, including those of the United States and its allies, use a
different signal processing architecture that includes cryptographic decoding and
increased accuracy.
Accuracy Obtained and Coordinate System
The two classes of GPS geolocation capabilities are the precise positioning service (PPS) and the standard positioning service (SPS). The PPS capability requires
cryptographic technology and achieves 22 m horizontal accuracy, 27.7 m vertical
accuracy, and 200 ns time accuracy. The SPS capability is available to any user
and achieves 100 m horizontal accuracy, 156-m vertical accuracy, and 340 ns time
accuracy. These specifications are defined by the 1999 Federal Radionavigation
Plan and are 95 percent accuracy figures (two standard deviations of radial error).
GPS Satellite Signals
GPS satellites transmit two spread spectrum signals, one at 1575.42 MHz (L1 for
SPS) and the other at 1227.60 MHz (L2 for PPS). There is a unique 1-MHz-wide,
1023-chip-long coarse acquisition pseudorandom (Gold code) spreading code for
each satellite [9]. A Gold code is a spreading code synthesized by exclusive
ORing of the output of two linear feedback shift register (LFSR) pseudorandom
(PN) generators together. Each LFSR uses a carefully selected tap and initialization to assure that the autocorrelations of the Gold code are small at all delays
except perfect time alignment, and are not confused with cross correlation from
other spreading codes.
272
Cognitive Techniques: Position Awareness
The coarse acquisition code on L1 is used for civilian GPS. There is also a
precision code (P-code), 10 MHz wide and repeating every 7 days, which is superimposed on L1 and L2. The P-code is used in precise mode. Encryption converts
the P-code into the Y-code. The P- and Y-codes are used for military GPS.
Additionally, there is a 50 Hz signal on the L1 coarse acquisition code that transmits such ancillary data as orbit parameters and clock data.
GPS Navigation Message
A data frame (1500 bits) is transmitted every 30 seconds and consists of five 300bit subframes.
●
●
●
●
●
Subframe 1 is Telemetry Word | Handover Word | Space Vehicle Clock
Correction Data.
Subframe 2 is Telemetry Word | Handover Word | Space Vehicle Ephemeris
Data part 1.
Subframe 3 is Telemetry Word | Handover Word | Space Vehicle Ephemeris
Data part 2.
Subframe 4 is Telemetry Word | Handover Word | Other Data.
Subframe 5 is Telemetry Word | Handover Word | Almanac Data for All Space
Vehicles.
Other data and almanac data are spread over 25 data frames and take 12.5 minutes to transmit. The 12.5-minute period is the complete navigation message.
Signal Processing of GPS Signals
The GPS receiver correlates known coarse acquisition spreading codes (with a
1-millisecond period of 1023 chips) from each of the GPS satellites with the
processed signal from the GPS satellites. The known spreading codes are very
short and may be generated or stored in memory. Because each satellite uses a different Gold-word spreading code, when the receiver has a peak correlation it
knows which satellite sent the signal. This despreading produces a full-power signal. This signal is tracked using a phase locked loop (PLL), and the 50 Hz navigation message is demodulated from each satellite. ToA information is extracted
when a correlation peak is measured.
Given the ToA information measured from the correlation peak and the GPS
time embedded in the signal, the GPS receiver can measure range to each satellite
in view. An intersection of multiple range spheres determines where the GPS
273
Chapter 8
receiver is located. Four satellites must be in view to estimate x, y, and z coordinates along with a time estimate. A precise estimate of the position of each space
vehicle in view is determined from the broadcast ephemeris data.
Reference Axes
The x, y, and z estimates are computed in Earth-centered fixed (ECF) coordinates.
ECF is a right-hand orthogonal Cartesian coordinate system with the origin at the
center of Earth, the z-axis increasing through the rotational North Pole of Earth,
the x-axis increasing through the prime meridian (Greenwich, England) at latitude
zero and longitude zero, and the y-axis increasing through 90 degrees longitude
and zero degrees latitude.
Differential GPS
Position accuracy may be improved through the use of differential GPS processing. Correcting bias errors using a known location accomplishes this. A known
location receiver measures its position and calculates a correction for each satellite that is passed to other GPS receivers in the local area. This is a sophisticated
solution that requires more capability at both the reference and mobile GPS
receiver and a data link between the reference receiver and the mobile receiver.
Another form of differential GPS is the measurement of carrier phase. This
capability is used in surveying and can generate subfoot accuracies over short distances. Again, at least two receivers, a reference and a mobile, are required. Due
to ionospheric effects, the set of receivers doing carrier phase geolocation must be
relatively close together (approximately a 30-km limit).
Potential GPS position error sources are summarized in Table 8.1.
Table 8.1: GPS errors. The intentional degradation from selective availability is a
significant source of position error in GPS systems.
Error source
Approximate error (m)
Noise (PN-code and receiver)
Selective availability*
Uncorrected clock errors (in space vehicle clocks)
Ephemeris data errors
Tropospheric delays
Unmodeled ionosphere delays
Multipath
2
100
1
1
1
10
0.5
*Selective availability is now off.
274
Cognitive Techniques: Position Awareness
8.2.2 Coordinate System Transformations
Satellite positions and GPS receiver position are reported in ECF coordinates
(x, y, z). Navigators, however, are frequently interested in latitude, longitude, and
height. Eq. (8.1) is the conversion from ECF to latitude, longitude, and height
(, , and h, respectively):
Latitude
Longitude
Height
⎛ Z e 2 b sin3 u ⎞⎟
⎟
f tan1 ⎜⎜⎜
⎜⎝P e 2 a cos3 u⎠⎟⎟
a
N (f) 1 e 2 sin 2 f
⎛ y⎞
l tan1 ⎜⎜ ⎟⎟⎟
⎜⎝ x ⎟⎠
P
h
N (f)
cos f
(8.1)
where
P x 2 y2
a semimajor Earth axis (ellipso id equatorial radius)
b semiminor Earth axis (ellipsoid polar radius)
⎛ za ⎞
u tan1 ⎜⎜ ⎟⎟⎟
⎜⎝Pb⎟⎠
a2 b2
.
e2 b2
8.2.3 GPS Geolocation Summary
A GPS receiver in a CR is one way to let a CR know where it is. Adding this
information to an inter-radio data stream enables other CRs to know where a particular radio is located. Even though GPS is not tremendously precise without differential mode or precision mode modules, it is good enough for policy enabling
and relatively long-range (similar to cell phone to base station) communication
optimization.
275
Chapter 8
8.3 Network Localization
8.3.1 Spatially Variant Network Service Availability
Geolocation of the radio (subscriber or user) units enables a number of useful
services. There may, for example, be services that the user wishes to find in his
immediate area. Consider the international traveler who has just landed after a
long flight from a foreign country and who is unfamiliar with the airport. He may
have several immediate needs including changing his currency, finding a restroom,
finding food, finding a power plug adapter, and finding a train to the final destination. If he is not fluent in the local language, his radio’s geolocation may be able
to give him the relative coordinates to find such services nearby, and perhaps even
help him negotiate the correct elevators.
Such processes require the ability to geolocate the user and then the requested
services. The requested services may come with caveats such as “the closest,” “the
least expensive,” or “near the route or path to be followed to some other service or
objective.” Having located the user, the system needs to be able to access local
networks and perform inquiries about available services, their respective locations, and then sort to the user’s preference criteria for presentation.
The above examples do not involve unsolicited advertising. However, there
may be a way for a CR to accept unsolicited advertising, filter it against a list of
topics of interest to the user, and present only those advertisements that pass the
interest filter. Those advertised services that do pass the interest filter test can be
offered to the user, prioritized with the user’s other objectives and any route planning so as to make them convenient.
Mechanization of service offerings becomes the major issue to enable this
degree of cognitive service based on geolocation. Assuming the CR has geolocated itself with one or more of the techniques described above, it needs to find a
gateway to local networks. If the CR has several personal, local, regional, and cellular network access points available to it, it may then query among these networks to find those that offer cognitive support to the types of queries that support
the user.
Assume, for the moment, that the CR has found an access point on a local network that offers such query services. Then the two radios may refine the location
of the user to sufficient accuracy to provide appropriate directions to find the
requested services. Those directions should be scaled by a scale factor appropriate
to the distance as well as the local knowledge and experience of the user. Once the
radio finds a service access point with the proper query service, the user may in
turn be directed to a different access point where that specific type of knowledge
276
Cognitive Techniques: Position Awareness
is served. If necessary, it may include a change to whichever wireless access point
is used to access the required service, as directed by the access process defined
from the first server.
Finally, notice that this database may be built in part by the success of other
users who have found useful services and captured the location of those services
into the database. Thus, the query engine may be able to build a significant library
of services without a significant system initialization effort.
Example 8.1 _______________________________________________
A student is connected through his WLAN to a campus-wide server that answers a
query about a restaurant with a special musical offering this evening, and the time
of the date. When seeking a new restaurant in an otherwise familiar city, the student need only know to go to a nearby street location, and then be given the final
guidance of where that restaurant is to be found (e.g., “The stairs to the second
floor restaurant are around the back of the clothing store on the first floor, and you
must go around the left side of the store to get to them.”). However, when the student is off campus, but near the restaurant, the CR switches to regional service, by
which the restaurant has a local query responder that can provide an applet to help
find the restaurant. The CR requires 3-D positional knowledge, as well as 3-D
navigation driving or walking, and, of course, the ability to understand the positional uncertainty of the user, the user’s current orientation and velocity, and the
relative position between the user and the objective. It is likely that local navigation may consist of applets offered by the endpoint server, in order to provide for
an endless supply of unique specialized access requirements (e.g., “take the last
elevator because it’s the only one that leads to the top floor”).
Example 8.2 _______________________________________________
Another example of geolocation-based services is emergency alert messaging.
Suppose that all users in a geographic region must be alerted to imminent danger.
Radios that know their position can be alerted, and the priority of all tasks in that
region can be completely changed. It may be necessary to provide such messages
over multiple networks, so there is a high probability that all CR radio users will
receive the message. It is also feasible to determine how many users remain in an
area and what their locations are, to expedite any remaining evacuation.
277
Chapter 8
8.3.2 Geolocation-Enabled Routing
In addition to the user’s geolocation support, the radio network functionality may
benefit from geolocation knowledge. Chapter 11 provides a detailed analysis of
the radio environment map (REM), which is an infrastructure to enable cognitive
network functionality.
However, in addition to the REM, the routing functionality of an ad hoc network,
and the cellular handover of cellular networks, may be improved by explicit geolocation knowledge, velocity vector knowledge, and planned route path knowledge.
One can envision that an ad hoc network could use destination location
addressing rather than medium access control (MAC) and IP addressing. In such
a network, messages propagate only to nodes that know they are on a path to that
location. It is expected that by reducing the number of nodes that are off course
for a destination, the end power drain from all nodes and the radio interference
level can be reduced. Cellular networks can improve their handover process to
anticipate handovers and reduce dropouts and dead zones by being aware of
geolocation, velocity vector, and planned path.
8.3.3 Miscellaneous Functions
A number of niche applications have been proposed that would be enabled by
CRs with geolocation knowledge, including radios that allow users to:
1.
2.
3.
4.
flirt or reject flirtation;
recognize when someone nearby is either desirable to meet or should be avoided;
recognize individuals with common or special interests who are nearby;
identify individuals with limited time budgets who cannot support any extraneous interactions; and
5. determine the location of criminals who have been released after serving time,
but who may still pose a threat under certain conditions.
This list is likely to grow as the economics of CR applications begin to benefit
users.
8.4 Additional Geolocation Approaches
Terrestrial radio geolocation is accomplished through one or more of the following techniques: ToA, TDoA, angle of arrival (AOA), or received signal strength
(RSS). Normally, a strong LOS signal is needed for accurate measurements.
278
Cognitive Techniques: Position Awareness
8.4.1 Time-Based Approaches
Time-based approaches for geolocation may be divided into ToA and TDoA
approaches. Both approaches require a high-resolution system clock. ToA and
TDoA approaches to geolocation are based on the propagation speed of light,
which is defined to be 3 108 m s1, and “straight line” LOS propagation paths so
that the signal time delay relates directly to the LOS distance. Propagation is analyzed by multiple receivers or by multiple antennas converting time delay to phase
difference, or time difference, or frequency shift. These are then translated into
equations of range, range ratio, AOA, and/or other parameters. Most systems also
track not only the estimated value but the error bounds, so the location can also be
represented as a circular error probability (CEP) ellipse. At the propagation speed
of light, 1 ns of time measurement error creates 0.3 m of ranging error.
ToA Approach
The ToA approach is centered on the ability to time-tag a transmitted signal and
measure the exact ToA of that signal at a receiver. The propagation time, at the
speed of light assuming LOS propagation, is a direct measure of the propagation
distance. This provides a receiver with an iso-range sphere for a given transmitted
and received signal. If multiple receivers at known locations receive the same signal, generally at different times, the multiple iso-range spheres intersect at the
transmitter’s location. It requires four receivers to geolocate one transmitter in
three dimensions. A reverse problem may be constructed in which four transmitters provide their location in a time-tagged transmitted signal and the receiver can
geolocate itself. More complex situations for which only relative positions can be
determined can be constructed. This problem is useful in sensor fields and also
for large numbers of CRs. A two-dimensional (2-D) depiction of this process is
shown in Figure 8.1.
Round-Trip Timing and Distance-Measuring Equipment
A valuable capability is being able to locate a transmitter, for example, to rescue
a sinking ship or a wrecked aircraft. Some aircraft are equipped with distancemeasuring equipment (DME) transponders. These systems respond to a transmission by transmitting a response pulse whose timing is precise relative to detecting
an arriving signal. By measuring the round-trip time of the interrogation pulse and
its arriving response, and then subtracting the response time of the transceiver’s
receiver–detector–transmitter, the total distance between the two radios can be
estimated. If the search transponder also uses three or more antennas, it can
279
Chapter 8
Figure 8.1: 2-D ToA: The
intersection of iso-range
spheres (3-D) or iso-range
circles (2-D) may not be
a point, introducing a CEP
(Rx: receiver; Tx: transmitter).
T3
Tx
Tx
Tx
T1
T2
Estimated Rx
Location with
CEP
estimate both the AOA and the distance to the distressed ship or aircraft relative
to its current location and can immediately fly directly to it or otherwise dispatch
aid directly to it. This has resulted in such products as the General Dynamics
PRC-112 transceiver and its interrogator.
Certain cellular radio systems have developed a similar DME approach to subscriber localization. In these systems, the cellular subscriber responds to transmissions from each of several cellular base stations of known position. Each cellular
base station can then estimate the range to that subscriber. By combining the
range estimates from each base station, the subscriber can be accurately located.
This method eliminates the requirement to install a separate GPS receiver function
in each subscriber unit and shifts the location complexity to the base station, thus
minimizing subscriber unit cost.
TDoA Approach2
The TDoA approach is centered on the ability to measure the time difference
between the reception of a signal at one location and the time of reception of the
same signal at another location. Again, propagation time, at the speed of light,
2
This section on Time Difference of Arrival Approach was contributed by Nicholas W. Fette,
July 2005, personal communication.
280
Cognitive Techniques: Position Awareness
is assumed to provide a direct measure of the LOS propagation distance. The
constant difference in arrival time produces a hyperboloid surface with the foci at
the two receivers. If multiple pairs of receivers at known locations receive the
same signal, the multiple iso-range hyperboloids intersect at the transmitter’s
location. It requires three surfaces to geolocate one transmitter in a 2-D plane, and
it requires four surfaces to geolocate one transmitter in three dimensions.
Fitting a TDoA Curve with Two Receivers
Suppose two radio transmitter–receivers (named 1 and 2), separated by a known
distance D1,2 in meters, each record the ToA of a signal from a source (named S) in
an unknown location, indicated by time stamps t1 and t2 in nanoseconds. Assuming
the signal propagated uniformly at the speed of light, c 0.300 m ns1, then the
speed of propagation multiplied by the difference between the values of these
time stamps is the difference of the distances from the source to the two radio
transmitter–receivers:
c(t1 t2 ) ct1,2 D1,S D2,S
(8.2)
This is one form of an equation for a hyperbola, the graph of which shows the
possible locations of S based on the two radio receivers. If both receivers and the
source are taken to be in the same plane, the hyperbola is roughly V-shaped
(Figure 8.2), with an axis of symmetry through the two receivers.
The hyperbola can be written in terms of a coordinate system (x, y) with
radio 1 (R1) at (0, 1⁄2 D1,2) and radio 2 (R2) at (0, 1⁄2 D1,2). A familiar form for a
hyperbola with these foci is:
y2 a 2 x2 k 2
(8.3)
with the constraints that y, a, and k carry the same sign as t1,2. The constants a
and k can be determined by inspection and are dependent only on D1,2 and ct1,2.
Consider the end behavior of the hyperbola. The signal from a hypothetical
distant transmitter (named HT) placed on the hyperbola at (xHT, yHT) for some
infinitely high xHT value travels in parallel paths through radios 1 and 2 and the
origin. A right triangle can then be formed (Figure 8.3) with the hypotenuse as the
segment from radio 1 (R1) to radio 2 (R2), one leg as a segment of length ct1,2
along the path to R1, and the second leg closing the right triangle at R2. Let be
281
Chapter 8
y
y 2a 2x2k 2
R2
k
u
x
ct1,2
R1
Figure 8.2: Hyperbola with two receivers and source in the same plane.
y
R2
u
u
D1,2
x
ct1,2
R1
Figure 8.3: Right triangle in the coordinate system (x, y).
282
Cognitive Techniques: Position Awareness
the angle between the hypotenuse and the second leg. By trigonometry,
sin Ct1,2/D1,2, and defining F Ct1,2/D1,2, the value of is:
arcsin f
(8.4)
By geometry, is also the angle from the x-axis to the infinite-distance transmitter, so tan y x . Thus, limx→yx a, a constant such that:
a tan tan(arcsin f ) f
1 f 2
(8.5)
whereas constant a is evaluated by considering infinite x, k is evaluated by moving the hypothetical transmitter to (0, k). Since HT is in line with radios 1 and 2,
D1,2 D1,HT D2,HT. Thus, D1,HT 1/2D1,2 k and D2,HT 1/2D1,2 k. By
Eq. (8.2):
ct1,2 D1,S D2,S 2 k
or
k
1
2
ct1,2
(8.6)
In summary, the hyperbola in this system is:
2
c 2 t1,2
2
y 2
1 c 2 t1,2
x2 1
4
2
c 2 t1,2
(8.7)
where y carries the sign of t1,2.
Transforming to a Common Coordinate System
The usefulness of the coordinate system used so far in Section 8.4 is limited to
communication between two radios only. To locate the signal source, at least three
radios must be employed. Therefore, all pair-wise hyperbolas must be transformed into a common coordinate system to solve for the source location. That
system may be chosen arbitrarily such that it enables conversion to and from GPS
coordinates recognized by each radio set; however, the latitude–longitude system
is warped and adds difficulty to the algebra in solving for the source location.
For example, the common coordinate system (x, y) could be oriented with y
directed north, x directed east, and the origin (0, 0) at the location of R1, and it
would form a valid system if the curvature of Earth is negligible, given fairly even
terrain. To transform Eq. (8.7) to this form, a rotation followed by a translation
283
Chapter 8
would suffice. The counterclockwise angle of rotation is determined by the
angle formed by the ray from R1 through R2 and the ray from R1 pointing northward (Figure 8.4).
Figure 8.4: Geometry for determining
counterclockwise angle of rotation
. This process would be used to
convert the hyperbolic equations in
radio-centered coordinate system
to a north centric coordinate system.
y
y
ψ
R2
x
R1
x
In this example, tan x1,2/y1,2, and the old temporary axes convert by:
x x cos y sin y y cos x sin (8.8)
where these substitutions are made into Eq. (8.3), in the manner:
(a 2 sin 2 cos 2 ) y 2 (a 2 sin 2 sin 2 ) xy (sin 2 a 2 cos 2 ) x 2 k 2 0
(8.9)
The solution of a system of at least two distinct equations in the form of Eq. (8.9)
can determine the possible position(s) (x, y) of S, the source of the transmission.
Solving for Position of Source Transmitter
The number of solutions that can result from systems of equations such as Eq. (8.9)
depends on the arrangement of the receivers and which pairs of receivers were
fitted with hyperbolas. As a rule, if all the receivers are in a line, there will be two
solutions, one on either side of the line. If all the receivers are on a single plane,
then there may be two height solutions, one of which may be underground.
284
Cognitive Techniques: Position Awareness
LORAN
LORAN systems transmit a known burst signal from multiple transmitters with a
known and published periodicity. Furthermore, the exact location of each transmitter is known. Three such transmitters cooperate to enable TDoA measurements. Ships at sea receive these transmissions and measure the time difference
between each received signal. From these time differences, ships are able to calculate the TDoA hyperbolas. To simplify the process, the TDoA hyperbolas have
been converted into published charts so that a navigator can directly look up the
time differences for each transmitter pair and find an intersection of two time difference pairs to perform a location at sea.
TV Broadcast
Recently, Rosum Corporation (Mountain View, California) announced that they
had developed a technique to recognize the ghost canceling reference (GCR) signal chirp that is included on the 19th line of the vertical retrace [9, 10]. By measuring the exact time that this chirp is received for several TV stations, and
comparing the time delay between the measured arrival and the precalculated
transmit time for that pulse, they can estimate a ToA for each detectable TV signal. TV transmitter locations are known with high accuracy, and they transmit
with high power, so this TDoA system can work in urban areas where satellite
signals fail. Since the bandwidth is 4.5 MHz versus the 1 MHz bandwidth of the
standard GPS signal, it may also improve spatial resolution by a factor of approximately 4. Finally, because networks include precision time information as a data
component during vertical retrace, all the information necessary for a sophisticated TDoA system is present in urban areas. The GCR chirp described above can
be used to recognize and suppress multipath components, and therefore can be
used in urban canyons.
Timing Estimates
Time can be derived from a number of sources, including atomic clocks, standard
clocks, GPS time, disciplined GPS, phase estimation techniques, and correlation
techniques in wideband transmission environments. The accuracy of the time estimate
directly influences the ranging estimate accuracy at approximately 1 ns 30 cm1.
A popular way to obtain TDoA information is through cross correlation. At a
fixed point in time (maybe on the GPS one pulse per second boundary), two stations digitize their received signals and pass the measurements to a common
285
Chapter 8
processing location. The processor calculates the cross correlation of the two signals. The peak of the cross correlation reveals the TDoA.
8.4.2 AOA Approach
The AOA approach requires an antenna array at the receivers. Multiple receivers
estimate the AOA of a signal. Combining the bearing to the signal with the known
location of multiple receivers yields an intersection point of the transmitter. This
is simple triangulation.
Geometry of AOA Approach
Geometry, of course, affects how well any time or angle measurements work,
including the AOA approach (Figure 8.5).3 If the “baseline” of receivers does not
have a sufficiently large angle of observation, the result is poor. If the accuracy
of the AOA sensors is poor and the range to the receiver is long, the “angular dispersion” impacts the measurement. However, the computations are simple if the
AOA is known to high accuracy and the interprocessor communications data
volume is low.
Figure 8.5: AOA: The geometry
affects the accuracy of AOA-based
geolocation estimates (Rx: receiver;
Tx: transmitter).
Tx
Estimated Rx
Location with
CEP
Tx
Tx
3
Geometric dilution of precision (GDOP) is one way of relating approximately the ranging errors to
position errors. Even though the distance measurement to each transmitter and its corresponding time
may be measured quite accurately, the geometry of the position of the transmitter as observed by the
receiver, may find that the lines of possible position intersect at shallow angles. This will cause the
intersection to have much larger positional error than the error from the range estimate alone.
286
Cognitive Techniques: Position Awareness
VHF VOR
Aircraft navigate using a number of transmitters located around the globe, frequently located near airports. These transmitters transmit an amplitude modulated
(AM) signal with a constant frequency. In addition, they transmit this signal
through a series of many antennas, each selected sequentially around a circle of
radius approximately 5 m. The energy of the transmission is thus electronically
swept around in a circle such that it introduces frequency modulated (FM)
Doppler onto the received signal. Each receiver will perceive the relative phase
angle between AM and FM as a different phase, reflecting the relative location
of the receiver relative to the transmitter. VOR receivers convert this modulation
phase angle into the angle bearing to the VOR. By locating two such VOR transmitters, the receiver is able to estimate its location.
8.4.3 RSS Approach
If the transmit power on a signal is accurately known, the patterns of the transmit
and receive antenna gains are known accurately, and the receiver is able to measure RSS accurately (all generally difficult assumptions to assure), then a propagation model can be used to estimate the distance to the transmitter as a function of
RSS. But propagation channels are very dynamic, so this approach is problematic.
The greatest source of error in this approach is multipath fading and shadowing,
and the effect can easily be a 40 dB impairment or a 6 dB increment over the
direct path LOS in as little as a half of a wavelength. Mobility allows a receiver
to average out these effects to some degree for FM music, but multipath represents
a significant impairment to data services.
This location approach is similar to the ToA approach. Four estimates of range
(upper and lower bounds of range) to four known transmitter locations are used to
calculate the intersection of iso-range spheres. This yields an estimate of position
and position error in three dimensions.
Simple propagation channel models predict path loss attenuation as an exponent ranging from 2 to 4.5, depending on terrain, foliage, and building shadowing
and blockage losses. Since the physical environment is such a key component to
other more realistic channel models, some averaging over position is required and
some setting of the attenuation exponent must be assumed. The setting may be a
table lookup as a function of rural, suburban, or urban environments and frequency.
If a correlation process based on a licensed transmitter’s database is undertaken, an RSS-based receiver application could determine in which regulatory
region it is located. For example, if a CR is receiving certain TV channels and
287
Chapter 8
certain FM stations and certain AM stations all at the same time, it can infer its
city location. If the location of the transmitters is included in the database along
with transmission levels, the RSS process could improve this estimate due to the
fairly large number of measurements.
The quality of RSS-based geolocation estimates is fairly low. It is useful to
CRs for some applications but not for others. For example, it may be close enough
to regionalize an estimate to a city or a part of a city, and thus to a regional accuracy of 30 km.
8.5 Network-Based Approaches
Wired networks have already developed a database that allows a table lookup
translation from IP addresses to geolocation. This database provides to the subscriber component the immediate ability to determine a geolocation to the accuracy of a service region. For WLAN devices, this allows the device to know its
location to within a radius of 100 m or so. Services offered from longer-range connectivity, such as fiber-optic or wired service, may be able to infer regional positional accuracy from such a table lookup. At the present time, the database does
not track its geographic uncertainty, nor does the database guarantee that it is currently up to date. Chapter 11 provides more information on how infrastructuresupported data services can provide geopositional support and timing to wireless
subscribers.
8.6 Boundary Decisions
The ability to determine the location of a CR to a geographic region enables at
least one beneficial capability. The benefit of automatic compliance with spatially
variant regulations has been discussed extensively in the CR community. Other
applications may develop from this capability as well.
8.6.1 Regulatory Region Selection
The capability of a CR to determine in which geopolitical region it is operating
enables spatially variant policy selection, and therefore worldwide mobility and
radio compliance with local regulatory rules. One approach for this capability is a
map database that contains a boundary set for each regulatory region. A hierarchal
search starting with continent, focusing into country (i.e., regulatory authority),
and finally selecting a specific policy region would minimize the time and computational load required to determine in which region a radio is located. The regulatory
288
Cognitive Techniques: Position Awareness
region returned is then used to select a set of policies. If the radio does not have
the correct policy, it may acquire it from locally available infrastructure, such as a
policy database server, or apply a default set applicable to the largest number of
regions in the last known continent.
Border Database Representation Analysis
This section explores several methods to determine the applicable geographic
region, and therefore the associated geopolitical region. Various methods are feasible for defining a geopolitical region based on current GPS coordinates of a
radio node. The objective is to find methods that minimize memory resources to
represent the regions a radio will experience during its product lifetime or mission
lifetime, so that it can choose corresponding policies for the associated region.
We assume in this analysis that the usual preference is for the mechanism that
consumes the least memory resource to represent each country’s borders. (There
are currently 192 countries in the world [11].) Specifically, we explore three methods to determine whether the current GPS location is inside a given complex polygon defining the borders of the region:
A. Successive tiling using latitude and longitude (east–west and north–south)
boundaries, in which the aggregated tiles define the geopolitical region.
B. A list of endpoints of successive line segments defining the borders.
C. A set of K nearest neighbor (KNN) position points.
Method A
In method A, we define a large rectangular region in the center of a country, and
then smaller rectangular tiles to fill in the shape of the irregular border regions. If
we assume one central region and an average of 100 smaller tiles, and four coordinates per tile, then each country would require, on average, A 101 4 404
points.
Searching this database can be quite efficient, particularly if we assume an outer
constraint rectangle as well. So, for example, if Figure 8.6 defines the maximum
northern and southern extent as well as the maximum eastern and western extent of
each country, then we can perform a subset analysis for any given GPS point that
can determine within which countries it may possibly lie. This is likely to include a
maximum of six countries. We then search the interior subtiles for each of those
possible countries, performing east, west, north, and south boundary comparisons,
289
Chapter 8
An estimated 360
point meets and
bounds description
for each region is
suggested
Each country may
further subdivide
regions by a factor of
10 where necessary
or convenient
Regions adjacent to
other countries
follow both sets of
regulatory policies
Figure 8.6: Defining geopolitical regions.
and stop as soon as we determine that a point is within a specific subtile. The resulting subtile is then associated to a country and the country to a policy set.
Method B
In method B, countries may define their borders as a linked list of straight-line
segments using GPS coordinates. We assume that sufficient resolution requires
a maximum of 360 2-D GPS locations, plus a country geographic centroid.
Therefore, B 2 361 722 points.
Although it may appear that this method is more frugal than method A in data
that must be stored, the search problem is more complex because the search algorithm must determine whether a point is enclosed inside or outside the region
defined by many line segments. To perform this calculation, we can begin, as in
method A, by determining which countries must be searched by comparing the
putative GPS coordinate against the extrema of all countries and only searching
those that are relevant. Again, we assume a maximum search size of six countries.
Next we calculate the perpendicular vector to each line segment. Along that vector, we calculate the minimum distance from the point to the line, and we save the
identity of the two closest line segments.
290
Cognitive Techniques: Position Awareness
For the sake of a simple explanation, we generalize that there are very few
instances of spiral borders. We can then determine whether most of the country
lies to the east, west, north, or south of the two line segments, and whether the
putative point is likewise to the east, west, north, or south of the two line segments. Failing that, we use the borders of each country until we find a case in
which the radio node location lies in the same direction as the centroid of the
country relative to the two nearest line segments. Note that this is equivalent to
dividing each country into a set of 100 triangular (pie-shaped) regions and determining whether a point is within the triangle.
Method C
In method C, we define a KNN approach to determine whether a point is in a
country. In this case, all countries must define a set of points approximately equidistant from the border that result in a KNN border that smoothly approximates
the true border.
We assume that the countries agree to place their KNN points at 10-km distance from the border and at approximately 10-km spacings. Furthermore, to
assure fairness, these points are placed on each side of the border along the same
perpendicular bisectors of border boundary line segments. This has the result of
smoothing fine-grain border detail to approximately 3-km feature sizes. The database size is now proportional to the length of the border. If we assume 100 sample
points for the average border, we must store C 100 2 200 points.
Furthermore, let us assume that K 2 in our KNN search. Now, for each
country, we calculate the Euclidean distance to all KNN locations, and we save
the final distance as the average of the two smallest distances. We compute this
calculation for all countries, and select the correct country as the one with the
smallest Euclidean distance. As with the other search processes, the search computation can be reduced to a subset of local neighbors by selecting to search only
those countries whose maximal extents cover the range where our point of interest
is currently located.
Anomalies
There may be countries that have anomalous shapes that may not behave correctly
when using the algorithms presented in section Border Database Representation
Analysis, such as spiral arms brought about by the circuitous path of rivers through
mountain ranges. In addition, islands are anomalies, but an island can be handled
as a second country with the same name and same policy.
291
Chapter 8
More likely sources of trouble, however, are disagreements about what the
actual border really is, such as overlapped regions claimed by more than one
country. Methods A and B can be extended to report that two countries both satisfy the criteria, defining the location of a radio. Under these conditions, it may be
possible to define a common subset policy or to report to the user that there are no
applicable rules in disputed zones.
The format for the latitude and longitude is negotiable. One suggestion is a
32-bit floating point number (4 bytes) representing degrees, minutes, and seconds
of latitude or longitude. This would result in a worst-case resolution along the
equator of about 5 m.
For each of the representation databases, and the conservative assumption of
200 countries, we can estimate the size of the database to be as shown in Table 8.2.
Note that at any one time only a subset of such a database is needed and the meets
and bounds database can be loaded into a CR similar to loading a working set into
a cache memory.
Table 8.2: Database sizes for three representation methods
to define whether a position is inside of a country.
Representation type
A
B
C
Database size (bytes)
323,200
577,600
160,000
8.6.2 Policy Servers and Regions
Possible sources of policy are: (1) a periodic broadcast (this protocol is not
defined at this time); (2) a load from a certified source (this could be realized in
the form of kiosks); or (3) a wired database server, which the device must proactively acquire, interrogate, and capture.
A policy region is not a defined entity at this time. An example of a policy
region might be the greater Washington, DC, area. In a specific regulatory region,
legacy transmitters are licensed, thus defining other broadcast bands that could be
harvested (only under specific technical circumstances). Another example of
something defined in a regulatory region is a licensed spectrum holder and a registry of frequency sublease server(s) (frequency and protocol) that may be contacted for leasing a channel for a short period of time. Both of these examples are
292
Cognitive Techniques: Position Awareness
“reach-out” capabilities and have not been approved in the United States. These
types of technologies have been discussed only in limited circles (see, e.g.,
Chapter 2).
In the short term, a small set of policy regions could be defined as boxes in
space and a subset of a small experimental band (there have been suggestions for an
experimental CR band) where demonstrations are conducted. Field demonstrations
of these capabilities could show the degree of spatial and frequency band reuse
achievable. Demonstrations also will show the ability of CRs to share the space on
a non-interference basis and the ability to comply with machine-readable policies.
8.6.3 Other Uses of Boundary Decisions
Radio policy selection is the frontrunner of CR applications that are dependent on
knowing in which geopolitical region a radio is located. Other potential applications also exist. Given natural language processing capability, a CR that knows it
is in Germany might offer to translate an English-speaking user’s word from
English into German when transmitting and could translate German words into
English when receiving. Although this could be accomplished through language
recognition more efficiently than through political boundary decisions, a boundary
decision could be used to improve the language recognition search by prioritizing
by local languages.
If a tourist is using the CR, the boundary decision engine could be used to display local points of interest such as historical markers. On a more personal note,
the boundary decision engine could be correlated with the user’s address book to
locate friends and family in the vicinity, and the user can apply this information
appropriately.
8.7 Example of Cellular Telephone 911 Geolocation for First Responders
The productivity of ubiquitous cellular telephone and handheld communications
devices such as pagers or BlackBerry™-type devices have dramatically increased.
Additional problems, however, have been introduced. For example, with wired
telephones emergency calls were traditionally easily located to the phone where
the call was placed and the location was very reliable. At the time of the initial
deployment of cellular telephones, a call placed to an emergency response center
did not come with a location attached for the emergency operator. The problem
prompted the US Federal Communications Commission (FCC), in 1996, to mandate geolocation services in the cellular network infrastructure. The FCC mandate
was for 125 m accuracy in 67 percent of all measurements by October 31, 2001.
293
Chapter 8
Multiple cellular telephone interfaces exist including Advanced Mobile Phone
System (AMPS), code division multiple access (CDMA) EIA/TIA IS-95, time
division multiple access (TDMA) EIA/TIA IS-136, the older TDMA IS-54, and
the Global System for Mobile Communications (GSM). The common characteristic of these systems is a set of base stations that communicate directly with mobile
stations or handsets. The frequencies, bandwidths, modulations, and protocols
vary from standard to standard.
The two broad categories of geolocation techniques for cellular telephones
are: (1) network- or infrastructure-based approaches and (2) handset-based
approaches. The advantage of the network-based approach is there are no requirements placed on the owners of the cell phones. The advantage of the handsetbased approach is the precision available.
Infrastructure-based approaches include ToA, TDoA, and AOA. The characteristics and techniques for these approaches have been discussed in Section 8.4.1.
The geolocation application executes on cooperating base stations, measures one
or more of the essential physical properties, exchanges information, and processes
the signals to produce a location estimate.
The best example of a handset-based approach is putting a GPS receiver into
the handset and interrogating the handset for its location when an emergency call
is placed.
8.8 Interfaces to Other Cognitive Technologies
This section discusses interfaces between the geolocation engine and other entities
in a CR. The policy engine, networking functions, planning engines, and user, at
least, will interface with the geolocation engine.
8.8.1 Interface to Policy Engines
One possible interface of the geolocation engine (server) and the policy engine
(client) is a client–server model. For example, the client requests from the server a
location in terms of latitude, longitude, and altitude. Another request would be for
a coded geopolitical regulatory region. These two requests would spawn different
signal processing steps to answer the inquiry.
Another interface to the policy engine is an “interference analysis” request.
This requires the relative location of other users to process the “message.” The
policy engine may request relative position directly to estimate the possibility of
interference. However, this does not address hidden node issues that are inherent
in the harvesting of spectrum.
294
Cognitive Techniques: Position Awareness
8.8.2 Interface to Networking Functions
The networking functions may use a similar client–server model to request relative position of other CR and non-CR devices. The resulting information may be
used to direct steerable antenna beams and nulls or to select next-hop destinations
in energy-efficient routing algorithms.
Networking functions may also request absolute position and use this in a
search for local services, such as access points. The database of services is contained in the networking engine in this case. If the database of services is contained in the geolocation engine, the networking function could request location
of the closest service provider, and the engine would return that. The partitioning
of these functions to the two engines has not yet been discussed in the literature.
8.8.3 Interface to Planning Engines
A new capability for the geolocation engine is a distance, as the crow flies, from
the CR’s current location to a designated position. An address to coordinate system function could be included. A variation of this request is from position A to
position B. The positions would need to be provided in a variety of coordinate
systems including ECF, latitude–longitude, or addresses. A planning engine executing a traveling salesman algorithm uses the geolocation engine for metrics. The
distances returned could be straight line or driving distances and paths, which may
be obtained from the Internet.
8.8.4 Interface to User
A user interface to the geolocation engine may be a simple “where am I?” that is
used to select a digital map segment. The map is overlaid with a “you are here”
marker. If relative positions of other emitters are requested, their positions may be
overlaid as well. This capability is very useful for navigation applications and
may be integrated with time and space management functions.
8.9 Summary
A CR that is aware of its position is enabled to demonstrate spatially variant behavior. This is a critical capability for new functions, such as spatially variant regulatory policy or spatially variant networking functions. If the radio knows where it is
located, it can self-report for such important functions as 911 responder interfaces.
Sections 8.2 and 8.4 covered a number of ways to determine where a system is
located. An inertial navigation system can be used to integrate a current position
295
Chapter 8
relative to a known starting point, but this approach is fraught with unreliability
and excessive expense. A better alternative is a GPS receiver. This inexpensive
subsystem provides 3-D position and current time. The system is based on a constellation of satellites and it has the capability for two resolutions, precise positioning system and standard positioning system. Very precise GPS location can be
obtained using differential approaches.
Non-infrastructure-based approaches include ToA, TDoA, AOA, and even
RSS. The first two approaches use spatial diversity and inter-radio communications to estimate a position for an emitter. A CEP may be calculated a priori and
is accurate with the exceptions introduced by high multipath channels. The AOA
approach uses spatial diversity and inter-radio communications to make a geometric interpretation of the location of an emitter. The last approach, RSS, yields a
regional estimate of position, but also uses spatial diversity and inter-radio communication to complete its estimate.
Section 8.3 touched on the value of geolocation knowledge to enable spatially
aware networking functions. Routing may take into account position for energy
savings or other purposes. Services may be accessed as a function of location.
New novelty functions, such as spatially variant advertising, is possible when a
CR knows its location.
Section 8.5 briefly looked at how networking can provide a coarse degree of
localization, and Section 8.6 explored how geolocation is extended to become
boundary analysis. One of the key interfaces to a geolocation engine is a boundary
decision. A meets and bounds database may be employed for a CR to obtain its
current geopolitical region. This approach may be used for a variety of other
regions of interest decisions as well.
Section 8.7 provided examples of geolocation in the context of cellular emergency location, and Section 8.8 reviewed the many interfaces to other supporting
subsystems a CR will need to build well-integrated, systems-level functionality.
These interfaces will be the areas for significant standards development as CR
technology evolves.
References
[1]
[2]
[3]
[4]
[5]
http://www.colorado.edu/geography/gcraft/notes/gps/gps_f.htm
http://tf.nist.gov/timefreq/general/pdf/1383.pdf
http://tf.nist.gov/timefreq/stations/wwvb.htm
http://tf.nist.gov/timefreq/stations/wwvh.htm
http://www.navfltsm.addr.com/vor-nav.htm
296
Cognitive Techniques: Position Awareness
[6]
[7]
[8]
[9]
[10]
http://www.lucent.com/press/0699/990630.bla.html
http://www.linuxjournal.com/article/7856
http://www.catchoday.com/archives/39006.html
http://www.uspto.gov patent search on patent #6,879,286
R. Gold, “Optimal Binary Sequences for Spread Spectrum Multiplexing,” IEEE
Transactions on Information Theory IT-13, 1967, pp. 619–621.
[11] www.google.com
297
This page intentionally left blank
CHAPTER 9
Cognitive Techniques:
Network Awareness
Jonathan M. Smith
CIS Department, University of Pennsylvania
Philadelphia, PA, USA
9.1 Introduction
Users see the network through the window of distributed applications, which
carry out some combination of communication and computation to meet user
needs. Familiar examples include web applications for shopping and banking,
interactive applications such as video teleconferencing, Voice over Internet
Protocol (VoIP) and chat, as well as music sharing and massive multiplayer
games. What each of these applications has in common is a model of the interactions between the user and the system, and how the system supports interaction
among sets of users.
The variety of applications listed above suggests that these models might
differ substantially, but in the abstract these models share the need for a set of
communications protocols that deliver network services required by the applications (Figure 9.1). Section 9.2 discusses specific requirements of typical applications, but as the understanding of networks and the variety of services they can
provide has improved with experience, the commonality among many application
requirements has been exploited to build network protocols that can be used by
many applications. The general-purpose nature (“one size fits all”) of such protocols is extremely attractive, but may sacrifice the ability to support specific applications as well as they might be supported by a dedicated and optimized protocol.
Note: Approved for public release, distribution unlimited.
299
Chapter 9
Figure 9.1: Protocols provide interfaces
and services to applications.
Applications
Protocols
Networks
9.2 Applications and their Requirements
There is an increasing trend toward mobile applications in the commercial world,
but mobility has always been a requirement for the military. Unlike those of the
commercial world, military communications challenges include lack of predeployed infrastructure, long latencies, higher bit error rates (BERs) than commercial
software systems assume, and higher standards for security. For example, the scenario illustrated in Figure 9.2 depicts a mobile “reachback” scenario, in which a
mobile satellite terminal is used to request data over a wireless satellite communications channel. Such a method may be slow and certainly represents substantial
propagation delay. If a conventional protocol architecture such as Hyper Text
Transfer Protocol (HTTP) 1.0 running over Transmission Control Protocol with
Internet Protocol (TCP/IP) is deployed end-to-end (i.e., between the mobile satellite terminal and the continental US ground data source) then the situation illustrated in Figure 9.3 can occur.
What occurs is that the application makes a request for data, such as an image
of a map, and the request is forwarded over the satellite channel. The data source
responds with the map, which can be substantial in size and will certainly require
multiple packets for transport. The application requirements here are reliable delivery and minimal latency. When TCP is used to obtain a reliable byte stream to
support the reliable delivery requirement, it also provides congestion control, a
feature desired by the network operator but generally not by an individual user.1
The TCP congestion control scheme interprets packet loss as indicative of congestion and therefore paces the number of packets in flight during a round-trip time
1
TCP provides two services: flow control, which ensures that the sender does not transmit more
packets than the receiver can handle, and congestion control, which mitigates the effect of an overflowed buffer in intermediate devices. The basic interaction between these two services is that the
flow control variable forms an upper bound to the size of the congestion control window.
300
Cognitive Techniques: Network Awareness
Figure 9.2: Reachback over a high bandwidth * delay product network.
Available Network Capacity
Bits per Second
Figure 9.3: Conceptual illustration of how
TCP congestion control adapts to packet
loss.
TCP – Reliable, But
Timeouts, “Slow Start”
Latency (Second)
Done
(RTT) (the “window size”) according to its perception of whether the network is
congested or not. Thus, even when capacity is available, as indicated by the horizontal dashed line in Figure 9.3, the conventional protocol architecture cannot
exploit it.
301
Chapter 9
Figure 9.4: Optimal network behavior
from the application perspective.
Bits per Second
Available Network Capacity
Full Capacity Reliable
Request/Response –
Minimal Delay
Latency (Second)
Done
Without saying more at this point in this chapter, Figure 9.4 should tease the
reader into thinking that something better is possible, the nature of which we will
explore in the following sections.
9.3 Network Solutions to Requirements
As shown in Section 9.2, there appears to be considerable opportunity to improve
the performance of applications if more is known about application requirements
and this knowledge can be brought to bear in the network. Before pursuing this, it
is worthwhile to examine the engineering issues at play in the protocol architecture implementations at work. Again, a convenient example for this analysis is the
widely used IP suite, which is often represented conceptually as an “hourglass” of
layers, as shown in Figure 9.5, where the hourglass is meant to indicate that the
interoperability layer, here labeled IP, is used to coerce multiple incompatible subnets into a “virtual” network, the inter-network [1], whose properties can then be
assumed to be present by end-to-end protocols. The inter-network is overlaid on
multiple sub-networks, and routes through the inter-network are used to provide
end-to-end connectivity to protocols such as TCP, which, as mentioned above,
provides a virtual-circuit-like reliable byte-stream model. HTTP is shown as a
user of TCP.
This architectural model has consequences for network awareness. The layering model employed reflects the software engineering mindset of the time: “modular programming” and “information hiding.” The idea of modular programming
is that more understandable (and therefore more robust) software can be written
when the software structure is decomposed into modules of small size, with localized concerns, such as a certain class of operation or a particular data object. Information hiding is a discipline that limits the flow of information among modules,
302
Cognitive Techniques: Network Awareness
HTTP
Network Service Overlay Protocols
Delays (from TCP Algorithms)
Losses (Generic; Unknown Cause)
TCP
UDP
IP
Losses (Probably Congestion)
ATM
Losses (Probably Error Bursts)
802.11b
Lightweight Virtual Infrastructure
(Global Packet Format, Addressing)
Subnetworks (Links Encapsulate IP)
SONET
Figure 9.5: The IP hourglass—an interoperability solution (ATM: asynchronous transfer
mode; SONET: synchronous optical network; UDP: User Datagram Protocol).
for example, by limiting data sharing to a small number of carefully type-checked
parameters.
In the case examined here, the concerns are localized in the layers, and the
information hiding discipline is enforced by the implementation, which passes
information either in the data structure for representing the packets (e.g., the memory buffer (mbuf) in UNIX implementations) or in the limited additional information used in procedure calls between layers. It is clear that such a discipline
enhances interoperability, a major goal achieved by the IP architecture. However,
like most architectural choices, it represents a particular choice in a rich space
of trade-offs, and as a consequence may introduce some undesirable properties.
As one example, we can consider the issue of packet loss. In many cases, if all
the information available to the end host was employed, an intelligent statistical
model for causes of packet loss could be employed. For example, the failure of
link layer checksums computed in the device driver for the particular line card
employed to connect the host to the network would be indicative of error bursts on
the link, whereas a recent trend of increasing delays in communication to a particular host or set of hosts might be indicative of congestion.
However, the IP protocol does not indicate why packets are lost (as above, it
could be burst errors, either causing whole packet loss or damage to a checksum,
or it could be congestion), so TCP must make an estimate with no information. On
commercial networks, the reasonable assumption as to the cause of loss is congestion. Thus, TCP slows down (sends less, in the decision matrix of Figure 9.6),
using a sophisticated algorithm that is clocked in units of RTTs.2 If, in fact, the
2
Because everything is clocked in RTTs, TCP’s “discovery” cycle is much slower in long-latency
networks such as those that incorporate satellite communications links.
303
Chapter 9
cause of loss is not congestion but some other cause, as is more common in
mobile military communications systems found in tactical scenarios, then havoc
ensues and performance plummets.
Figure 9.6: Matrix of loss causality
versus congestion control actions
(ARQ: automatic repeat request;
FEC: forward error correction).
Ideal
Behavior
Send
Less
Send
More
(FEC/
ARQ)
Congestion
Loss?
Error
Loss?
9.4 Coping with the Complex Trade-Space
One possible approach to overcoming certain limitations of information hiding, as
reflected in protocol architectures, is shown in Figure 9.7. The left-hand side illustrates that the limits are imposed by narrow service interfaces, as discussed in
Section 9.3. These service interfaces may be, for a variety of reasons, immutable
(e.g., because a large body of existing software assumes the interfaces and the
economic consequences of changing the interface would be unacceptable). In such
circumstances, a protocol architecture could be adjusted [2], as shown in the righthand side of Figure 9.7, by inserting a “shim” module into the protocol stack
which is compliant with the interfaces defined above and below it in the layered
architecture. In the example illustrated in Figure 9.7, an additional protocol function, automatic repeat request (ARQ), has been inserted beneath the IP layer to
Figure 9.7: Protocols built with
service interfaces. Service interfaces
are often narrow and opaque. In
spite of this, structural adaptation
can occur through dynamic protocol
composition.
Application
Application
Application
(Narrow)
Service Interface
TCP
TCP
IP
IP
Protocol
Link
304
ARQ
Learn
Link
Cognitive Techniques: Network Awareness
reduce the observed losses due to errors. The addition of this protocol function
reduces the probability of burst errors being observed by the IP layer. Thus, the
assumption by the TCP layer that packet loss is due to congestion is made more
probable, and the congestion control strategy is therefore much more likely to
respond correctly to the packet loss events that it observes post-forward error
correction (FEC).
Since the ARQ protocol seems useful, why not always employ it at the link
layer? This is, in fact, done for certain types of links. For example, the 802.11
medium access control (MAC) layer employs an ARQ-like strategy to emulate the
reliability of a wired local area network (LAN). Some satellite communications
links employ performance-enhancing proxies (PEPs) at the link endpoints to provide a reliable substrate to end-to-end protocols. Such link protocols are not universally employed, however, due to the performance consequences, for example,
increased latency (the 802.11 MAC may account for much more measured latency
than propagation delay). It should be clear, then, that an adaptive protocol architecture that employs the error compensation mechanism, as needed, is desirable.
In the illustration on the right-hand side of Figure 9.7, we might learn that there
has recently been an increase in the checksum failures on link layer frames, and
deploy ARQ or an FEC scheme. Over time, the wisdom of this decision can be
balanced against the cost of the deployment, likely duration of error bursts, hysteresis models for the module removal decision, and so forth. Section 9.6 further
discusses the robust use of system measurements on the Defense Advanced
Research Projects Agency (DARPA) SAPIENT program (see Section 9.6).
Generalizing from this example, we can see that structural adaptation of protocols [2, 3] is a powerful technique with which we can provide application performance in spite of an extremely wide range of application requirements and
network conditions, by adding and removing protocol elements in response to
changes in the situation. Although this problem might be solved in specific circumstances by event-driven adaptations, such as the insertion of ARQ discussed
in this section, there is a wide variety of additional information that must be available to fully exploit structural adaptation in network protocols.
Consider an application that requires data privacy and integrity. For such an
application, it may be desirable to employ an encryption algorithm, which protects data privacy through a secret-controlled data transformation, and as a consequence of this transformation, detects data tampering as a failure to successfully
reverse the transformation at the receiver. Redundancy present in the application
data and limited network throughput might suggest that a data compression protocol be employed.
305
Chapter 9
Consider the two possibilities for ordering compression and encryption protocol elements illustrated in Figure 9.8. In the compress/encrypt composition of two
protocol elements, as might occur in the course of a structural adaptation process
such as that discussed earlier, if compression is performed first, the data size is
reduced by removing redundancy, and then an encryption transformation is performed to hide the bits of the compressed data. If, in contrast, the encryption is
performed first, then the structure exploited by a compression algorithm is
obscured, meaning that no data size reduction will be accomplished in spite of the
considerable computation typically required for protocol elements of this type. So
the first composition is desirable and meets the goals of “compressed encrypted,”
whereas the second composition meets only the “encrypted” goal.
Figure 9.8: Protocol composition possibilities.
Protocol composition must be aware of
protocol element interactions.
Compress
Encrypt
Encrypt
Compress
This example illustrates the point that several kinds of knowledge must be
available in a general structural adaptation scheme. First is knowledge of what
external conditions the protocol element is responsive to, such as network conditions and application requirements. Second is knowledge of the assumptions
inherent in particular protocol elements responsive to external conditions, which
may affect its suitability and role in a complete protocol architecture resulting
from a multiplicity of structural adaptations. In the past, the daunting complexity
of managing the many possible adaptations and their possible interactions has
inhibited adaptations for which simple robust models (e.g., the control laws
employed for TCP/IP’s adaptation to network congestion, or the definition of a
“spanning tree” [3]) have not been devised.
9.5 Cognition to the Rescue
A new approach, which employs what can broadly be viewed as an approach to
automating the management of complexity, now seems possible. Automating the
management of complexity does not seek to eliminate or obscure the complexity
inherent in a system (as the information hiding approach might be seen to do).
306
Cognitive Techniques: Network Awareness
Application Performance (e.g.,
Goodput, Latency, Connectivity)
Rather, automation seeks to remove the detailed management of complexity from
the purview of the application programmer or protocol designer. Because many
protocol design issues are quite subtle, as has been illustrated in this discussion,
any such protocol automation strategy must be able to represent many kinds of
knowledge. Thus, empirical knowledge (such as observed performance) should be
able to be “fed back” into the knowledge employed by the system to create protocol structures suitable for particular applications and network conditions.
The opportunity afforded by the approach is considerable. Figure 9.9 illustrates the possibilities for structural adaptations that closely couple application
requirements to the encountered network conditions (which, of course, as in the
packet loss example, may vary considerably with time). When the network conditions are very attractive, as might be the case on the interconnection network of a
parallel computer or a set of hosts resident on a common LAN, very little protocol
support may be necessary for typical desiderata such as high “goodput” with low
latency. Where conditions degrade, more and more protocol mechanisms may be
employed to adapt the application to the degraded network, with these compensatory mechanisms permitting the application to operate in domains unreachable by
traditional architectures (the limitations of which are illustrated by the demarcation of the vertical dotted line in Figure 9.9).
Remove
Unused
High
St
an atic
d A St
lgo ruct
rith ure
ms
Situa
tionProto aware
cols
Add
Compensatory
Failure
of Static
Protocol
Low
Ideal
Network Conditions
Poor
Figure 9.9: Performance of network conditions. Situation-Aware Protocols occupy a new
region of a conceptual “trade-space.”
A detailed example of the possibilities can be given by using the TCP/IP congestion control algorithm’s response to packet loss. Simulations and laboratory
experiments have both shown that this protocol’s response to error conditions
above a certain threshold (e.g., a BER greater than 106) causes the TCP/IP algorithm to close its congestion window. As a consequence, its performance is
307
Chapter 9
reduced to the level of a “stop-and-wait” protocol, which is extremely sensitive to
the delay inherent in the link. If error compensation is inserted (as in Figure 9.7),
the windowing strategy is made effective, and relatively high throughput can be
achieved in operating regimes with high BERs. Hadzic and Smith [4], for example, have shown experimentally that a protocol stack incorporating TCP/IP and
adaptive compensation can operate with BERs as large as 104.
Cognitive systems are systems that apply human-style reasoning and capabilities. In the case of the structural adaptations discussed here, the cognitive system
would be required to identify the combination of application requirements and
network conditions, select or devise an appropriate composition of protocol elements, and deploy that composition. Notions such as similarity, trial-and-error,
and improvement by learning appropriate responses over time help to make a cognitive system both less brittle and more evolvable than a conventional protocol
design. It provides a clean global separation of policy and mechanism, which is
otherwise done only locally in modules, if at all.
9.6 The DARPA SAPIENT Program
The clear utility of such an approach in military networks, particularly those at the
tactical “edge,” has stimulated DARPA to support a research effort to investigate
the effectiveness of cognitive approaches to rapid adaptive composition and adaptation of protocol structures. This effort is called SAPIENT, for Situation-Aware
Protocols In Edge Network Technologies, and has identified three core challenges
that must be addressed in a working system:
1. Knowledge representation
2. Learning
3. Selection and composition.
Presuming that these challenges are addressed, a conceptual architecture for
an SAPIENT system might be structured similarly to that shown in Figure 9.10.
An application that employs the network to carry out its intended task (such as
those discussed in Section 9.2), is shown in the upper left of the illustration.
The requirements of the application might be characterized explicitly (by
a designer or user), although it is far more attractive if the SAPIENT system
deduces or infers these requirements automatically. An example of explicit characterization might be the quality of service (QoS) trade-offs specified in Nahrstedt
and Smith’s [5] OMEGA system, whereas automatic inference might be done by
308
Cognitive Techniques: Network Awareness
Application
Learn
Input or Infer
Protocol Element
Knowledge
Application
Requirements
Select/Compose
Protocol Elements
Sense Network
Conditions
Composed Protocol
Figure 9.10: One possible structure for a Situation-Aware Protocol System.
technologies such as statistical machine learning operating on the packet stream
emitted by the application.
As a simple example of automatic inference, consider observing packet size
and a time stamp. A relatively constant packet size and inter-arrival delay might
indicate the packets are part of a media stream, and the protocol support might be
adjusted to support this inference. Conversely, great variability in timing, small
sent packets, and highly variable sizes in received packets might indicate a transactional application such as that discussed at the end of Section 9.2 and the desirable protocol behavior illustrated in Figure 9.4.
The goal is to produce a knowledge base of application requirements, as
shown in the upper right side of Figure 9.10. This knowledge is intended to be
combined with other knowledge, namely, that about network conditions and available protocol elements, to produce a responsive SAPIENT system. The protocol
element knowledge, illustrated at the center of Figure 9.10, is used to react to network conditions. It is brought together with the application information in the
process of selecting and composing protocol elements, which results in the composed protocol illustrated in the lower right side of Figure 9.10.
Many opportunities for learning remain. Learning to automatically identify
applications and their requirements has already been discussed, but the sensing of
network conditions (losses, timing, load, etc.) is central to understanding whether
a new protocol is needed, and if so, what network conditions it must address.
Sensing network conditions can involve a variety of sensors, including those
that monitor packet checksums, packet ordering, packet sequencing, packet loss,
congestion indications, latencies, and relative application usage of link capacity,
309
Chapter 9
as well as alerts such as might occur on the loss of a link or a dramatic reduction
in signal strength on a wireless link. Sophisticated sensors might include hysteresis or damping algorithms to compensate for stochastic “noise” that might otherwise induce unwanted responses. Input from multiple low-level sensors might be
fused to gain a better overall estimate of the state of the network; such sensory
information could include information obtained through active probes as well as
from remote systems, whether they employ an SAPIENT architecture or not.
The sensing of network conditions is a central element of an SAPIENT system’s “situation awareness” because it provides the measurement basis for maintaining the current state or pursuing an adaptation. When the system is initialized,
it might start by using a composition equivalent to TCP/IP as a baseline, and then
modify the system while a positive performance gradient is realized. If a protocol
composition is not working, then the network sensing process should deduce this
and stimulate production of a more appropriate protocol. When a performance
“plateau” is achieved, the SAPIENT system would preserve the status quo until a
significant change in network conditions occurred.
Over time, the protocol element knowledge base incorporates this sensed
information in its learned responses to detected network conditions. Here, statistical machine learning can prove powerful, for example, in classifying situations in
which the operating protocol should be left untouched, and those in which it
should be adjusted. One property possible in such a system is the ability to
recover from protocol elements with errors, presuming that the sensed data is correct and that it is correctly interpreted.
The use of these varieties of knowledge in creating situation awareness, resulting in application-specialized protocols, offers a new way to support networked
applications.
9.7 Summary
This chapter has presented a new form of network awareness—situation awareness—and its application in the construction of application-specific protocols.
Use in the fashion described creates an evolvable system, as new protocol elements
and knowledge about them can be inserted directly into the system. This would
permit, for example, data from an “experienced” SAPIENT system to be loaded
into a new system, in essence sharing the learned knowledge. The structure illustrated in Figure 9.10 permits the system, after this initial load, to continue learning
as it encounters novel combinations of network situations local to its operating
environment.
310
Cognitive Techniques: Network Awareness
The SAPIENT effort and its approach to managing complexity represent a
pathfinder to new methods of building robust systems that can operate successfully in extreme environments and unforeseen conditions, such as those faced by
military tactical communications systems on a daily basis.
References
[1] V.G. Cerf and R.E. Kahn, “A Protocol for Packet Network Intercommunication,”
IEEE Transactions on Communications COM-22, May 1974, pp. 637–648.
[2] D.C. Feldmeier, A.J. McAuley, J.M. Smith, D.S. Bakin, W.S. Marcus and
T.M. Raleigh, “Protocol Boosters,” IEEE Journal on Selected Areas in
Communications, Vol. 16, No. 3, April 1998, pp. 437–444.
[3] D.S. Alexander, M.S. Shaw, S.M. Nettles and J.M. Smith, “Active Bridging,” in
Proceedings of the ACM SIGCOMM, Cannes, FR, 1997, pp. 101–111.
[4] I. Hadzic and J.M. Smith, “Balancing Performance and Flexibility with Hardware
support for Network Architectures,” ACM Transactions on Computer Systems,
Vol. 21, No. 4, November 2003, pp. 375–411.
[5] K. Nahrstedt and J.M. Smith, “Design, Implementations and Experiences of the
OMEGA Endpoint Architecture,” IEEE Journal on Selected Areas in
Communications, Vol. 14, No. 7, September 1996, pp. 1263–1279.
311
This page intentionally left blank
CHAPTER 10
Cognitive Services for the User
Joseph P. Campbell, William M. Campbell,
Scott M. Lewandowski and Clifford J. Weinstein
Information Systems Technology Group, MIT Lincoln Laboratory,
Lexington, MA, USA
10.1 Introduction
Software-defined cognitive radios (CRs) use voice as a primary input/output modality and are expected to have substantial computational resources capable of supporting advanced speech- and audio-processing applications. This chapter extends
previous work on military speech applications (see, e.g., Ref. [1]) to cognitive-like
services that enhance military mission capability by capitalizing on automatic
processes, such as speech-information extraction and understanding the environment.
Such capabilities go beyond interaction with the intended user of the softwaredefined radio (SDR)—they extend to speech and audio applications that can be
applied to information that has been extracted from voice and acoustic noise, gathered from other users and entities in the environment. For example, in a military environment, situational awareness and understanding could be enhanced by informing
users based on processing voice and noise from both friendly and hostile forces operating in a given battle space. This chapter provides a survey of a number of speechand audio-processing technologies and their potential applications to CR, including:
●
●
A description of the technology and its current state of practice.
An explanation of how the technology is currently being applied, or could be
applied, to CR.
Note: This work was sponsored by the Department of Defense under Air Force contract
FA8721-05-C-0002. Opinions, interpretations, conclusions, and recommendations are those of
the authors and are not necessarily endorsed by the US Government.
313
Chapter 10
●
●
Descriptions and concepts of operations for how the technology can be applied
to benefit users of CRs.
A description of relevant future research directions for both the speech and
audio technologies and their applications to CR.
A pictorial overview of many of the core technologies with some applications presented in the following sections is shown in Figure 10.1. Also shown are some
overlapping components between the technologies. For example, Gaussian mixture models (GMM) and support vector machines (SVM) are used in both speaker
and language recognition technologies [2]. These technologies and components
are described in further detail in the following sections of this chapter.
Speech and concierge cognitive services and their corresponding applications
are covered in the following sections. The services covered include speaker recognition, language identification (LID), text-to-speech (TTS) conversion, speech-totext (STT) conversion, machine translation (MT), background noise suppression,
speech coding, speaker characterization, noise management, noise characterization,
and concierge services. These technologies and their potential applications to CR are
discussed at varying levels of detail commensurate with their innovation and utility.
10.2 Speech and Language Processing
The following speech- and language-processing technologies begin with the
acoustic speech signal collected from a single microphone (multisensor collection
using a variety of sensors is also shown). All the speech and language technologies here can be viewed as being abstractly related to the anatomy and dynamics
of the vocal apparatus and behaviors expressed via speech and language and/or
viewed as statistical modeling methods.
10.2.1 Speaker Recognition
Speaker recognition can enable a CR to authenticate users for access control,
identify communicating parties, personalize the device, and adapt the device
and its applications to individuals. Speaker recognition technologies allow systems
to automatically determine who is talking or, to be precise, whether the incoming
voice compares favorably with an enrolled user’s voice. This determination can
then be used to provide user authentication for access control, identification of
communicating parties, and personalization and adaptation of the device and its
applications. Figure 10.2 shows the basic operations of a speaker recognition
system and its two phases of operation: enrollment and verification.
314
Speaker Recognition Evaluations:
Speech Enhancement: Modification of Time, Pitch, Spectrum
TD Pitch, Onset,
and Voicing
Estimates
Miss Probability (%)
40
Periodic
Component
20
10
5
Speech
In
2-Period Onset
Sync. Window
Pitch Sync.
Decomposition
2
1
0.5
0.2
0.1
Speech
Out
Pitch Sync.
OLA
Time -Scale
Modification
0.1 0.5 2
0.2
1
5 10 20
Residual
Component
Pitch
Modification
40
Enhancement and
Noise Suppression
False Alarm Probability (%)
Speaker/LanguageRecognition
Speaker/Language
RecognitionModels
Models
Speech
Coding
Gaussian Mixture Model
Speech Signal
Support Vector Machine
Enhanced Speech
Low -Rate Speech
600 – 2400 bits/sec
Speaker
Recognition
Speaker Name
John Q. Public
Language
Recognition
Language Name
English
V
315
Ngram Language Model
N-gram
Spectral
Modification
IY
G
Speech Recognition and
Machine Translation
Words
- are you? ”
“ How
Multisensor
Coding
EY
Res-Mic(s ) P-Mic
GEMS
New System Evaluation Techniques:
• Readability of speech-to-text output with metadata
• Comprehension of machine translation output
Language ID System:
PPRLM Core
Recognizer
GMM Core
Recognizer
SVM
SVM Core
Core
Recognizer
Recognizer
F
U
S
I
O
N
The colorful ceremony include parents and
friends were young wedding ceremony
Ibrahim Rashid Ahmed ; On the ; Miss
Eveline appeal ; The family of the magazine
make newlyweds my sincere congratulations
and wishes them happiness happiness.
Questions:
1. What is the purpose of this article?
2. What message does the magazine‘s staff add?
Multisensor
Enhancement
Signal Fusion
Noise/Speech
Preprocessing
10-point
Intelligibility
Gain in M2
Noise
Multisensor Coding
Figure 10.1: Speech and language technologies. A wide variety of speech and language technologies, such as speaker and
language recognition and MT can be used to enable cognitive-like services for software-defined radios. In addition,
multisensor speech coding and enhancement technologies can aid users, especially in harsh noise environments.
Chapter 10
Enrollment
Phase
Enrollment Speech
for Each Speaker
Bob
Model for Each
Speaker
Feature
Extraction
Model
Training
Bob
Sally
Enrolled Speaker Database
Verification
Phase
Feature
Extraction
Verification
Decision
Sally
Accepted!
Claimed Identity: Sally
Figure 10.2: Speaker recognition. Speaker identification is based on extracting features from
speech that best characterize the anatomical and behavioral differences of each individual
when compared to the distributions of those features among the entire speaker population.
Those differences are used to identify speakers or verify the claim of a speaker’s identity.
Enrollment and Verification
In the enrollment phase, voice samples from the subject are used to create a model
or template, which is sometimes improperly referred to as a voiceprint, for the
specific speaker. In the verification phase, the unknown voice is compared with
the model of the claimed identity. A verification decision can be made to accept or
reject the identity claim.
Speaker recognition is imperfect and is characterized by two types of errors:
false alarm (FA—meaning false recognition) and miss (failure to recognize the
claimed enrolled speaker). These systems are characterized by whether the speech
they use is text dependent (e.g., phrase prompted or pass phrases) or text independent (e.g., conversational speech). The performance of these systems is quantified by a plot of the miss rate versus the FA rate, which vary as a parametric
tradeoff of a control threshold selected by the system designer called the operating
point. The operating point is adjusted by varying an acceptance threshold that
shifts tradeoffs between the two error types up or down the curve of a given system. Often, a combined measure is cited to provide an approximate representation
316
Cognitive Services for the User
of overall system accuracy; this measure, known as the equal-error rate (EER),
indicates the operating point at which the miss and FA rates are equal. The stateof-the-art text-independent speaker recognition performance for conversational
telephone speech of a few minutes in duration is in the range of 7–12 percent EER
[3]. Given an extended duration of enrollment speech, even better performance
can be achieved, as shown in Figure 10.3.
Fuse10
Word-SVM
40
Cep-gmm
Cep-SVM
Cep-svm+nap
Prosodics
Phone-SVM
Pitch+bandeng-slope
Probability of Miss (%)
20
Pitch+eng-slope
Pitch+eng-gmm
Word-IIr
Word-IIr
10
Word-SVM
Cep
Phone-SVM
5
Phone-Ilr
Fuse10
2
1
0.5
0.2
0.1
0.1 0.2
0.5
1
2
5
10
20
40
False Alarm probaility (in %)
Figure 10.3: Performance curves of speaker recognition systems. Shown are the performance
tradeoffs between probability of miss and false alarm of ten algorithms and their fusion.
The speaker models were trained on approximately 20 minutes of speech and tested on
about 2 minutes of speech.
As introduced in Campbell et al. [4], voice biometrics1 are well suited to
radios that already incorporate microphones and speech coders2 (necessary for
encrypted voice communications). Additionally, voice biometrics can be combined with other biometrics, such as face recognition, as shown in Figure 10.4, for
increased security and/or backup operating modes (e.g., face recognition could be
more reliable than speaker recognition in high-noise environments or vice versa in
1
Biometrics is the automatic recognition of individuals based on biological and behavioral traits.
Low-rate speech-coding vocoders parameterize the speech signal for transmission. This parameterization can be shared with speaker recognition engines, typically with some loss in performance, to conserve processing.
2
317
Chapter 10
Machine Gun
Fire
Audio
Feature
Extraction
Babble
Video
Feature
Extraction
Classification/
Fusion
RF Sensors
(GEMS)
Feature
Extraction
Aircraft
Noise
Tank
Noise
Contact Sensors
(Pmic, Accelerometer)
Feature
Extraction
Speaker
Identity
Figure 10.4: Speaker identification combined with facial and other biometrics. This
combination enhances the accuracy, reliability, and survivability of the biometric system.
Additionally, the suite of sensors shown enables robust low-rate speech coding in extremely
harsh military noise environments [5].
adverse lighting conditions). Given the proliferation and improving quality of
cameras in cellular phones, CRs are also likely to have cameras that might be suitable for face recognition.
Voice biometrics can provide access control via biometric logins and screen
locks. This includes guarding against unauthorized use of lost CRs, such as disabling an idle radio that has been left behind. Voice biometrics can also enable
user conveniences, such as recalling preferences and adapting to users (as do
modern operating systems and application software). Conventional biometrics are
generalized here to incorporate an additional authentication factor by learning and
recognizing the users’ distinctive behaviors. Future research directions for speaker
recognition focus on making it more robust to mismatched channel conditions and
applying high-level features that resemble those used by humans [6–8].
User Authentication
This section presents various means of user authentication, introduces generalized
biometrics, and illustrates continuous and confidence-based user authentication.
As shown in Figure 10.5, the four pillars of user authentication are knowledge
(e.g., PIN or password), tokens (e.g., key or badge), behaviors (e.g., usage
318
Cognitive Services for the User
Strongest
Authentication
Behaviors
Traits
Do
Are
Know
Have
Knowledge
Tokens
Figure 10.5: The four pillars of user authentication. Recognizing people based on what they
do, what they know, who they are, and/or what they have. Using multiple attributes can
increase reliability and survivability. Using all four attributes can provide the strongest
authentication [4].
patterns or outcomes), and traits (e.g., voice or fingerprint). The proper combination of all four pillars provides the strongest user authentication. Biometrics are
used to authenticate users,3 as opposed to authenticating something they know
(such as a password, which can be forgotten or compromised) or possess (such as
an identification card, which can be lost or stolen). Unlike knowledge- and tokenbased authenticators, however, the inability of users to transfer biometrics can
lead to difficulties; as in, for example, an emergency transfer of operation of a
radio with biometric access control to an unenrolled user. To solve this difficulty,
knowledge- and token-based authenticators can be used to authenticate users in
these situations.
Popular biometrics includes voice, face, fingerprint, and iris.4 Voice and face
biometrics (possibly in combination) are well suited to radios that already incorporate microphones and cameras. Some biometrics lend themselves to continuous
user authentication (e.g., to guard against lost or stolen radios being misused) and
allow a system to assess varying levels of trust. For example, voice verification
can be used to continuously authenticate a user while the user is talking; this can
be useful if the voice quality makes it difficult for the interlocutor to detect a
change in talkers. Figure 10.6 shows an example of an authentication process over
time with varying levels of trust [4]. This example begins in a state of provisional
trust and, over time, proceeds in continued states of provisional trust and then to a
3
Strictly speaking, biometric verification is a binary hypothesis test. Here, the hypothesis is that
the live voice matches an enrolled and stored voice model. The biometric system decides to accept
or reject this hypothesis.
4
See http://www.biometrics.org/ for additional biometrics.
319
Chapter 10
Trusted State
Required for Sensitive Operations
Provisional Trust
Continue Interaction, Gather
Behavioral and Voice Samples
Trust
Time
Untrusted State
Interrupt Interaction
Figure 10.6: Continuous user authentication and trust. Varying levels of trust assessed over
time via continuous authentication of a user are represented as state transitions. With
sufficiently long samples of speech, the system is generally able to make a confident decision
about the speaker’s identity and, therefore, whether or not to trust the user [4].
trusted or untrusted state. While in a state of provisional trust, benign operations
(e.g., adjusting radio volume) can be performed, whereas sensitive operations
(e.g., downloading an SDR waveform) require a trusted state.
Behavior-based user authentication recognizes users via their actions, interests, tendencies, preferences, and other patterns. Examples of distinctive behaviors include:
1. How a user does something (e.g., speed and pattern of typing, stylus angle and
intensity, use of menus versus keyboard shortcuts).
2. What a user does (e.g., patterns of applications use, program features used,
patterns of collaboration).
3. What a user causes to happen (e.g., sequences of system calls, patterns of
resource access).
These behaviors include not only a user’s local actions, but also network interactions and outcomes. Behavior-based user authentication, like voice verification,
has minimal adverse impact on a user’s normal activities. The authentication is
inherent and transparent; there is continuous mode operation and modest resource
utilization, and user acceptance is likely to be high. Monitoring these behaviors
can be combined with situational awareness to fuse multiple factors into the
authentication process.
320
Cognitive Services for the User
A cognitive approach allows for many interesting possibilities. First, the
threshold to reach the trusted state of user authentication can be adapted based on
situational, environmental, and mission awareness and the risk of the requested
operation (e.g., benign volume adjustment to sensitive security operations). Second,
authentication can be performed over time by combining available information—
voice communication, mouse/stylus movement, dialogue structure, etc.
Some issues and questions in biometric deployments involve:
1. Whether to use remote versus distributed versus network enrollment and
verification.
2. Where user models are created and stored.
3. How models are maintained and updated.
4. How enrollment is conducted.
5. How models are bound to users.
6. What the tolerable verification time or rate is.
7. How models of new users are distributed and their integrity assured.
8. Whether there are accuracy or policy requirements.
9. What the architecture to support the biometrics is.
Biometric Sensors
There are many approaches to biometric-based user authentication and all require
some form of hardware input device to gather the required information about the
user to be authenticated. For example, fingerprint recognition requires a fingerprint scanner, user voice recognition requires a microphone, and user behavior
monitoring requires a traditional user input device (e.g., a keyboard or mouse).
These hardware devices must be an integral part of the CR platform and must
communicate with their software counterparts over a secure channel. Many of
these devices are high bandwidth, although the utilization is often in bursts. The
channel connecting the hardware and software must be capable of supporting the
data transfer requirements without an undue performance impact on the device’s
core functionality.
Security Architecture with Biometric Processing
Once data has been gathered from a biometric sensor, it must be processed to
determine user identity (or other user characteristics that the sensor has been
321
Chapter 10
designed to assess). Such processing can occur either in software, as shown in
Figure 10.7, or in specialized hardware. Although the use of specialized hardware
provides the advantages of increased tamper-proofing and higher performance, the
complexity of managing updates and modifications to the functionality of the
hardware often outweighs these benefits. Implementing the biometric processor in
software does not fundamentally diminish the overall security of the system, given
that the overall security of the radios (and the network services provided by
radios) is also built on software components. Although not shown in Figure 10.7,
most biometric processors require access to a database that contains information
required to authenticate users, such as biometric models or templates, profiles, or
logs of past behavior.
Figure 10.7: Notional radio
security architecture. This
architecture supports
continuous-biometric
processing and trust
assessment over time for CR
security [4].
Network
Encrypted
Comm
Authentication
Interface
Authentication
API
Security
Policy
Security
API
Security
Manager
Biometric
Processor
Biometric
Sensors
Operating System
Applications
Radio
Applications
Although the platform has been designed to accommodate existing applications
without modification, new applications can be designed to leverage the capabilities exposed by the security application programming interface to improve the
322
Cognitive Services for the User
user experience and to improve overall system performance. Designing applications to leverage this architecture is critical to its success [4].
Legacy applications will automatically benefit from continuous authentication,
but will be completely unaware of confidence-based authentication. Therefore, the
platform will need to define a confidence level at which the user is considered
authenticated, and any level of confidence below that will cause the user to be
considered completely unauthenticated by legacy applications. Applications that
are aware of confidence-based authentication, in contrast, can enable functionality
or access to data based on the confidence of the user’s identity.
10.2.2 Language Identification
LID can enable a CR to infer whether its user or a remote communicator is a
friend, a neutral party, or a foe based on identifying the language transmitted over
a radio channel. LID technologies allow systems to automatically determine the
language of the user from a list of possibilities. These technologies are available
for over a dozen languages. These systems usually require about 30 seconds of
speech to obtain good spoken LID performance.
Methods for spoken language recognition have traditionally been based on
phonetic transcription of different languages [9]. By discovering the relation between
occurrences of phones5 in different languages (i.e., phonotactics6), a statistical
model can be constructed of a particular language for later use in identification.
An emerging class of recent methods for language recognition is based on
novel features [10]. These new features, shifted-delta cepstral coefficients,7 measure changes in the speech spectrum over multiple frames8 of speech to model
long-term language characteristics. These methods need only a speech corpus
labeled by language for training in order to achieve good results.
As mentioned previously, current system performance [10] is measured in
terms of FA rate and target miss rate for detectors of individual languages. Stateof-the-art error rates for speech from telephone environments are shown in
5
Phones are subunits of the basic sound units of speech (e.g., “ah” or “t.”).
The set of allowed arrangements or sequences of speech sounds in a given language.
7
The cepstral coefficients are common features used in speech-signal processing that are derived
from a speech signal’s spectral content. The cepstrum, pronounced kepstrum and an anagram of
spectrum, is the result of taking the Fourier transform of the decibel spectrum, as if it were a signal.
The mel-frequency cepstrum coefficients are a common variation and popular features used in
speech and speaker recognition.
8
Each frame of speech is on the order of 20 milliseconds in duration.
6
323
Chapter 10
Figure 10.8. This plot shows average results for the following 12 languages:
Arabic, English, Farsi, French, German, Hindi, Japanese, Korean, Mandarin,
Spanish, Tamil, and Vietnamese. Results are shown for males (m), females (f),
and both males and females (b) speakers for 3-second, 10-second, and 30-second
utterances. EERs are less than 3 percent for 30-second utterances.
03s-m
03s-f
03s-b
10s-m
10s-f
10s-b
30s-m
30s-f
30s-b
40
Miss Probability (%)
20
10
5
2
1
0.5
0.2
0.1
0.1 0.2
0.5
1
2
5
10
20
40
False Alarm Probability (%)
Figure 10.8: Typical performance of a language recognition system. Results are taken from
the 2003 National Institute of Standards and Technology (NIST) Language Recognition
Evaluation. Performance improves (curves shift toward origin) as the test speech duration
increases and performance is fairly independent of the talker’s sex.
LID has many potential applications in CR. First, LID could be used as a
defense against system overrun; that is, the system could allow only certain languages to be used for radio communications. A more experimental strategy may
be to look for shibboleths9 to recognize the actual dialect or accent of the speaker;
for example, whether the speaker has a foreign accent. A second application of
LID is in situational awareness. If speech communications can be intercepted (and
9
This biblical term has come into modern usage as a linguistic test to determine members of one
group versus outsiders.
324
Cognitive Services for the User
decrypted), their language could be identified to aid in the recognition of friends
and foes and to alert soldiers of changes in the languages spoken by nearby
forces.
10.2.3 Text-to-Speech Conversion
TTS conversion can provide status information to an eyes-busy CR user. TTS
technology automatically speaks textual information. Textual information could
originate from text-based communications or equipment display readouts. Textbased examples of communications that could be spoken via TTS includes e-mail,
news, web, rich site summary (RSS), weblogs (blogs), instant messaging (IM),
Internet relay chat (IRC), and short message service (SMS). Traditional equipment display readouts could also be spoken via TTS, such as radio frequency, battery power, signal strength, network speed, time, speed, location, and bearing.
By providing status information to an eyes-busy user, TTS would enable soldiers to focus on their mission while hearing an explanation of their battle space
and status. Different synthesized voice types (e.g., male and female) are usually
employed to convey different types of information. For example, routine or urgent
information could be conveyed in male or female voices, respectively.
The current state of TTS technology produces mostly reasonable sounding
speech; however, it does not yet sound quite human. Future research directions in
TTS are focusing on improving the quality of voice synthesis, pronunciation of
named entities, conveyance of expression, and integration with MT and STT
conversion.
10.2.4 Speech-to-Text Conversion
STT conversion can recognize a CR user’s voice commands, take dictation, and
compose text messages from voice. STT technology attempts to automatically
convert speech into a form that can be read by a user, produce entire transcripts of
a conversation (continuous speech recognition), perform word spotting (looking
for particular words), and provide command and control functions via voice. As
CR keypads and displays shrink, the voice control modality becomes even more
useful.
Recently, speech recognition has developed along several paths. One path is
work on large vocabulary, continuous speech recognition for conversational situations. This work has been funded through projects such as the Defense Advanced
Research Projects Agency’s (DARPA) Effective Affordable Reusable Speech
Recognition (EARS) program; work in this area can be found in, for example,
325
Chapter 10
Schwartz et al. [11]. Progress in STT has brought error rates down to less than
12 percent word error rate for telephone speech. Another recent path for STT
work is in noise robustness. An overview of some of these methods can be found
in Junqua and Haton [12]. Noise robustness has been studied extensively for standardization by the European Telecommunications Standards Institute (ETSI) for
distributed speech recognition (DSR), as exemplified by Parihar and Picone [13].
DSR’s goal is to make STT a client–server application, in which the client uses
the DSR front-end to parameterize the speech, while recognition is done on the
server.
STT has many possible applications in CR:
First, STT can be used for gisting—rather than having a user listen to the complete conversation, a summarized version of the output can be produced.
Second, STT can be used to route certain conversations to appropriate users
(see Riccardi and Gorin [14] and related references).
Third, STT can be used for data mining speech. If radio communication is
processed by STT and stored, then text-retrieval techniques (such as those
used to search for documents on the Internet) can be a quick and efficient way
of searching for content.
Fourth, STT can be used for command and control of a CR, as described by
Brown and Campbell [15]. In this scenario, a speech interface frees up tactile
and visual modalities so that the user can more effectively multitask. The
speech interface can be used to control various aspects of the CR: radio
modes, sensor interfaces, sensor analysis, etc.
10.2.5 Machine Translation
MT automatically converts words or phrases from one language into another. This
is generally done on text; however, MT can be combined with STT and/or TTS
conversions to provide mixed mode translation [16].
MT technology could help a soldier during operations in foreign-language
environments. For example, foreign-language signs, news, and radio intercepts
could be roughly translated to the soldier’s language to aid in understanding the
battle space.
Current MT technology, as typified by various web-based systems, can be
helpful for extracting some of the key words and phrases from foreign-language
material, but such translations are by no means transparent, as they generally contain many errors. Transcription problems are frequent and are often, but not
326
Cognitive Services for the User
always, easily detectable by users (it could be argued that it is more problematic
when users are unable to detect transcription problems). Future MT-related
research will likely be aimed at improving basic MT performance, automatically
extracting meaning, gisting, and summarization.
10.2.6 Background Noise Suppression
Background noise suppression enhances communication in the presence of noise,
but extracts useful background noises to enhance situational awareness. Background noise suppression is primarily used in conjunction with STT conversion
and voice communication. For voice communication, many new technologies
have become available over the last few years.
Noise suppression can be used in voice communication to enhance the effectiveness of a speech coder. In this case, a noise suppression system attempts to
improve both the quality and the intelligibility of coded speech. These methods
fall into several categories. First, methods that attempt to “subtract out” the noise
spectrum have achieved considerable success (see, e.g., Ref. [17]). A sophisticated form of spectral subtraction [18, 19] is integrated in the enhanced mixedexcitation linear-predictive (MELPe) 1200/2400 bps speech coder [20]. A second
class of noise suppression algorithms is based on computational auditory scene
analysis (CASA) [21]. The idea, in this case, is to use algorithms inspired by
human processing; people effectively separate a sound field into multiple components such as music, voice, noise, among others. CASA methods use techniques
such as independent component analysis and array processing to achieve noise
suppression. A third class of noise suppression methods is based on multimodality. A well-known phenomenon for humans is that visual processing and audio
processing of speech are fused, as evidenced by the McGurk effect10 [22]. Several
systems have been developed to take advantage of the visual component (see, e.g.,
Ref. [23]). Alternate non-acoustic modalities have also been explored; these
include electromagnetic, accelerometer, and electroglottogram (EGG) sensors.
Significant improvement in noise suppression has been achieved with these
approaches [5].
Active noise suppression is another technology that is being incorporated into
radio systems. Active noise suppression reduces the noise that a user perceives by
emitting sound to cancel the undesired noise field. Active noise suppression can
10
The McGurk effect is a phenomenon that indicates an interaction between hearing and vision in
speech perception.
327
Chapter 10
be used to decrease fatigue caused by exposure to high noise levels and reduce the
Lombard effect.11 Consumer headphones incorporating active (and passive) noise
suppression are commercially available.
Noise suppression is a critical component of a CR with a speech user interface. Although not usually perceived as a cognitive capability, noise suppression
is ultimately a test of a system’s ability to deal with real-world conditions.
Techniques such as multimodality and CASA show the sophistication and the
challenge of matching human processing in this task.
10.2.7 Speech Coding
Speech coding is the effective use of varying or limited capacity channels for
voice communications, as is widely experienced via cellular telephony. Speech
coding seeks to provide communicability, intelligible speech, quality speech,
talker and state (e.g., stress) recognizability, low delay for two-way communications (the total system one-way delay must be under 300 milliseconds for normal
conversation), talker and language independence, naturalness, robustness in
acoustic noise (including background talkers), insensitivity to transmission errors,
tandem (synchronous and asynchronous) coding capability, ability to transmit
signaling/information tones, near minimum rate coding (variable bit rates can complicates encryption and conferencing), and minimal computational and memory
complexities to maximize battery life. Speech coding designers trade off many of
these objectives when designing or choosing from among the many coders for
real-world applications.
The current generation of speech coding standards capitalizes on the fact that
people listen to speech communications systems and, thus, the systems attempt to
minimize perceptual distortion. Future research will likely focus on optimizing
coding for speech, as opposed to other types of signals; taking advantage of the
language being spoken; widening the analysis bandwidths; and fusing multiple
sensor streams. New adaptive speech coders will not only provide good
communications-quality speech under noisy conditions, but will operate at
dramatically reduced bit rates to conserve battery life and/or provide high
processing gain to decrease the probability of a communication being intercepted,
detected, or jammed [24]. This will yield improved and safer voice communications for the soldier.
11
The Lombard effect is the tendency to increase one’s vocal effort, which alters speech, in noise.
328
Cognitive Services for the User
10.2.8 Speaker Stress Characterization
Speaker characterization is the process of automatically determining the state of a
user by using speech-processing techniques. Typically, this has meant trying to
determine a person’s emotional state, which is often related to the stress a user is
experiencing.
Speaker characterization is still a developing science. One of the difficulties is
elicitation of an emotional state for corpus collection—how can an experimenter
truly ensure that a participant is stressed and still comply with Institutional
Review Board human subject testing requirements? Another difficulty is the definition of emotional states. For example, stress can take many forms: physical
stress, emotional stress, task-based stress, noise-induced stress, and so on. Should
all of these be separate categories of stress? Regardless of the experimental difficulties, several practical techniques for stress recognition and compensation have
been examined; for examples, see the earlier work [25, 26] and work on the
Speech Under Simulated and Actual Stress (SUSAS) corpus [27].
Speaker characterization is related to CR in various ways. Speaker characterization can be part of a broader strategy of affective computing [28]. Some examples include:
●
●
●
●
Knowing the stress state of the local user as well as other users in the field to
improve situational awareness.
Knowing if a user is irritated by a particular feature by relying on the user’s
voice characteristics.
Using stress level to determine appropriate modality (e.g., visual versus audio)
for response to a query.
Using verbal cues to determine if the CR made a correct decision.
Ultimately, if a CR is able to perceive and exploit a user’s emotional state, it
will make better decisions and be a more effective device.
10.2.9 Noise Characterization
Although noise has been considered a nuisance up to this point, it can be a useful
source of information. Noise can be exploited in several ways. First, noise characterization can provide situational awareness; it can infer the location and situation
of friends and foes by characterizing the acoustic environment in which the radio is
operating. The CR, or a user, could catalog and track noise types in an environment
329
Chapter 10
to recognize anomalies that might indicate the presence of friend or foe. In this
case, a noise characterization system would have to find features and provide
recognition of different types of noise sources: vehicles, guns, planes, and so on.
Also, the directionality of noise sources in the acoustic and radio-frequency environments would be a critical property to assess. Second, noise characterization
can provide diagnostics. Noise analysis can be used to detect imminent mechanical failure of common military equipment [29] or provide a quick diagnosis of
mechanical problems.
10.3 Concierge Services
With CR, there is an opportunity to provide sophisticated human interfaces.
Broadly speaking, these interfaces fall into two different categories—transparent
and concierge. For a transparent interface, the goal is to provide services to the
user without the user explicitly invoking the interface. A simple example of this
service would be automatic detection of input modality. A more advanced interface example is an augmented reality interface where the user sees information
overlaid on real objects through a heads-up display. The other interface type,
concierge, requires the user to explicitly invoke and respond to cognitive services.
An example of this type of service is an agent that searches for information on a
topic based on a verbal request. Hybrid architectures between these two interfaces
are also a possibility—we may not want services automatically invoked, but we
may want the CR to transparently know the user.
A unique opportunity in CR is the combination of advanced interfaces with a
portable radio. A simple example that illustrates some of the ideas involved is
shown in Figure 10.9. Suppose two users are communicating, and that they want
to decide on a rendezvous for lunch (time, place). Both CRs have knowledge of
their users’ preferences and current status (money, schedule, etc.). Both radios
also have a common ontology for communicating about the users’ preferences.
When one user asks to have lunch, the two radios can communicate and reason
about a rendezvous place. This problem has many facets and can be quite sophisticated; the system will have to reason about:
●
●
●
Location: The location of the two users and how long it would take to go to a
common restaurant.
Schedule: The available schedules of the two users, and how traveling to lunch
would impact these schedules.
Resources: How much money do the two users have? Is the restaurant available?
330
Cognitive Services for the User
Preferred Restaurants: Thai,
Italian, …
Recent Choices: McDonald’s,
…
Money in wallet: $$$$
Where am I? 45 Canal Street
Schedule: open 1-2 pm
Preferred Restaurants:
Chinese, …
Recent Choices: Bertucci’s,
…
Money in wallet: $$
Where am I? 23 Main Street
Schedule: open 1:30-2:30
Cognitive Radio
Agent
Communication
Where’s a good
place?
What’s nearby?
What kind of food do
we like?
What time can we
meet?
Let’s get
lunch …
Figure 10.9: An example concierge service for a CR. The radios’ concierge services make
arrangements for a rendezvous for lunch. In a cognitive-like manner, the radio might
anticipate the users’ needs for lunch and sense nearby associates to invite.
●
●
●
●
●
User preferences: What type of restaurant is reasonable for both users? What
prior choices have the users made?
Social network: If they invite other people, how does this impact the choices?
Dynamic rescheduling: The CR could suggest an alternative rendezvous, if new
information was discovered en route to the first choice.
Network resources: The CR will access a comprehensive list of restaurants
stored on the web.
User intent: Is this meeting business or personal? Can the radio anticipate a
lunch date from the user’s schedule?
Ideally, the CR would find a list of prioritized restaurant choices. The users could
then verbally negotiate the remaining details.
The example highlights the need for several technologies for cognitive interfaces. First, a common method of describing the user state and preferences is
needed. For many situations, a language such as Web Ontology Language (OWL)
would be appropriate. A second technology of importance would be modeling
user intent. Several options exist in this area—Hidden Markov Models or
Dynamic Bayesian Network models, for example Ref. [30] and [31], and
plan/goal recognition [32]. These technologies attempt to predict what task the
331
Chapter 10
user is trying to accomplish and the current state of the task. Third, technologies
that mimic a human interface may be appropriate—avatars, speech technologies (as
previously discussed), visual interfaces, haptic interfaces, etc. Finally, sophisticated
reasoning technologies must be able to handle time, geography, and uncertainty.
Reasoning and learning in a CR to support the user interface consists of many
different types; three significant examples are reactive, adaptive, and networkbased. Reactive reasoning is quick and does not involve explicit planning. A cognitive system would be expected to respond immediately to user voice commands,
user preferences, etc.; that is, a reactive reasoner would be local to the CR. An
adaptive CR would learn from its past experiences by adapting user preferences,
speech recognition models, user intent models, etc. It could also pass this knowledge on to other CRs. Finally, one of the large potentials for CR would be in
network-based reasoning. A CR could access a large network of services to
respond to user requests. For instance, in the concierge example, the agent
representing the user would be expected to obtain information about location,
restaurants, travel time, etc. The benefit of network-based reasoning is flexibility
and minimization of local resource usage; the drawback is latency and protocol
standardization (a common communication ontology).
10.4 Summary
This chapter has provided an overview of several speech- and audio-processing
technologies exploiting the voice and acoustic noise streams that are likely to be
available to a software-defined CR.
An integrated approach to user authentication and architecture to enhance
trusted radio communications networks has been presented. User authentication,
via generalized biometrics, can be combined with other authenticators to provide
a continuous, flexible, and strong system. This biometrically enhanced authentication system approach can be extended to become part of a CR system that learns
about users, situations, and surroundings and takes appropriate proactive or reactive actions. One kind of learning presented here was generalized biometric
authentication, where the users’ distinctive behaviors and traits are learned and
recognized. Cognitive-like applications to CR were given using speaker recognition, LID, TTS conversion, STT conversion, MT, background noise suppression,
speech coding, speaker characterization, noise management, noise characterization, and concierge services technologies. An advanced CR will capitalize on
these technologies to learn about and take action based on user preferences, availability of resources, and other elements of the situation and environment.
332
Cognitive Services for the User
These technologies leverage the significant computational capabilities of
future CRs to improve the capability, effectiveness, and efficiency of users of CRs.
A high-impact implication of this technology is the ability to drastically reduce
the complexity of the man-machine interface with the radio. As these technologies
mature and become more robust, they will provide significant efficiency for the
user, which will better enable the user to accomplish his most important tasks.
References
[1] C.J. Weinstein, “Opportunities for Advanced Speech Processing in Military
Computer-Based Systems,” Proceedings of the IEEE, Vol. 79, No. 11, 1991,
pp. 1626–1641.
[2] W.M. Campbell, J.P. Campbell, D.A. Reynolds, E. Singer and P.A. TorresCarrasquillo, “Support Vector Machines for Speaker and Language Recognition,”
Computer Speech and Language, Vol. 20, No. 2–3, 2006, pp. 210–229.
[3] M.A. Przybocki and A.F. Martin, “NIST Speaker Recognition Evaluation
Chronicles,” in Proceedings of Odyssey: The Speaker and Language Recognition
Workshop, Toledo, Spain, 2004, pp. 15–22.
[4] J.P. Campbell, W.M. Campbell, D.A. Jones, S.M. Lewandowski, D.A. Reynolds
and C.J. Weinstein, “Biometrically Enhanced Software-Defined Radios,” in
Proceedings of Software Defined Radio Technical Conference (SDR ’03), Orlando,
Florida, 2003.
[5] T.F. Quatieri, K. Brady, D. Messing, J.P. Campbell, W.M. Campbell,
M.S. Brandstein, C.J. Weinstein, J.D. Tardelli and P.D. Gatewood, “Exploiting
Nonacoustic Sensors for Speech Encoding,” IEEE Transactions on Audio,
Speech, and Language Processing, Vol. 14, No. 2, 2006, pp. 533–544.
[6] J.P. Campbell, D.A. Reynolds and R.B. Dunn, “Fusing High- and Low-Level
Features for Speaker Recognition,” in Proceedings of Eurospeech, Geneva,
Switzerland, 2003, pp. 2665–2668.
[7] W.M. Campbell, J.P. Campbell, D.A. Reynolds, D.A. Jones and T.R. Leek,
“High-Level Speaker Verification with Support Vector Machines,” in Proceedings
of International Conference on Acoustics, Speech, and Signal Processing, Hong
Kong, China, 2003, pp. 73–76.
[8] W.M. Campbell, “Compensating for Mismatch in High-Level Speaker
Recognition,” to appear in Proceedings of IEEE Odyssey: The Speaker and
Language Recognition Workshop, San Juan, Puerto Rico, 2006.
[9] M.A. Zissman, “Comparison of Four Approaches to Automatic Language
Identification of Telephone Speech,” IEEE Transactions on Speech and Audio
Processing, Vol. 4, No. 1, 1996, pp. 31–44.
333
Chapter 10
[10] E. Singer, P.A. Torres-Carrasquillo, T.P. Gleason, W.M. Campbell and
D.A. Reynolds, “Acoustic, Phonetic, and Discriminative Approaches to Automatic
Language Recognition,” in Proceedings of Eurospeech, Geneva, Switzerland, 2003,
pp. 1345–1348.
[11] R. Schwartz, T. Colthurst, H. Gish, R. Iyer, C.-L. Kao, D. Liu, O. Kimball,
J. Makhoul, S. Matsouka, L. Nguyen, M. Noamany, R. Prasad, B. Xiang, D. Xu,
J.-L. Gauvain, L. Lamel, H. Schwenk, G. Adda, L. Chen and J. Ma, “Speech
Recognition in Multiple Languages and Domains: The 2003 BBN/LIMSI EARS
System,” in Proceedings of International Conference on Acoustics, Speech, and
Signal Processing, Montréal, Québec, Canada, 2004, pp. 753–756.
[12] J.-C. Junqua and J.-P. Haton, Robustness in Automatic Speech Recognition, Kluwer
Academic Publishers, Boston, Massachusetts, 1996.
[13] N. Parihar and J. Picone, “Analysis of the Aurora Large Vocabulary Evaluations,” in
Proceedings of Eurospeech, Geneva, Switzerland, 2003, pp. 337–340.
[14] G. Riccardi and A.L. Gorin, “Stochastic Language Adaptation over Time and State
in Natural Spoken Dialog Systems,” IEEE Transactions on Speech and Audio
Processing, Vol. 8, No. 1, 2000, pp. 3–10.
[15] C. Broun and W.M. Campbell, “Force XXI Land Warrior: A Systems Approach to
Speech Recognition,” in Proceedings of International Conference on Acoustics,
Speech, and Signal Processing, Salt Lake City, Utah, 2001, pp. 973–976.
[16] W. Shen, B. Delaney and T.R. Anderson, “The MIT-LL/AFRL MT System,” in
Proceedings of International Workshop on Spoken Language Translation,
Pittsburgh, Pennsylvania, 2005.
[17] Y. Ephraim and D. Malah, “Speech Enhancement Using a Minimum Mean-Square
Error Log-Spectral Amplitude Estimator,” IEEE Transactions on Acoustics, Speech,
and Signal Processing, Vol. 33, No. 2, 1985, pp. 443–445.
[18] R. Martin, I. Wittke and P. Jax, “Optimized Estimation of Spectral Parameters for
the Coding of Noisy Speech,” in Proceedings of International Conference on
Acoustics, Speech, and Signal Processing, Istanbul, Turkey, 2000, pp. 1479–1482.
[19] R. Martin, D. Malah, R.V. Cox and A.J. Accardi, “A Noise Reduction Preprocessor
for Mobile Voice Communication,” EURASIP Journal on Applied Signal
Processing, Vol. 2004, No. 8, 2004, pp. 1046–1058.
[20] T. Wang, K. Koishida, V. Cuperman, A. Gersho and J.S. Collura, “A 1200/2400 bps
Coding Suite Based on MELP,” in Proceedings of 2002 IEEE Workshop on Speech
Coding, Tsukuba, Ibaraki, Japan, 2002, pp. 90–92.
[21] M. Cooke and D. Ellis, “The Auditory Organization of Speech and Other Sources
in Listeners and Computational Models,” Speech Communication, Vol. 35, No. 3–4,
2001, pp. 141–177.
[22] H. McGurk and J. MacDonald, “Hearing Lips and Seeing Voices,” in Nature,
Vol. 264, 1976, pp. 746–748.
334
Cognitive Services for the User
[23] X. Zhang, C.C. Broun, R.M. Mersereau and M.A. Clements, “Automatic
Speechreading with Applications to Human-Computer Interfaces,” EURASIP
Journal on Applied Signal Processing, Vol. 2002, No. 11, 2002, pp. 1228–1247.
[24] A. McCree, “A Scalable Phonetic Vocoder Framework Using Joint Predictive
Vector Quantization of MELP Parameters,” in Proceedings of International
Conference on Acoustics, Speech, and Signal Processing, Toulouse, France, 2006.
[25] K. Cummings and M.A. Clements, “Analysis of the Glottal Excitation of
Emotionally Stressed Speech,” Journal of the Acoustical Society of America,
Vol. 98, No. 1, 1995, pp. 88–98.
[26] J. Hansen and M. Clements, “Source Generation Equalization and Enhancement of
Spectral Properties for Robust Speech Recognition in Noise and Stress,” IEEE
Transactions of Speech and Audio Processing, Vol. 3, No. 5, 1995, pp. 407–415.
[27] S.E. Bou-Ghazale and J.H.L. Hansen, “Speech Feature Modeling for Robust
Stressed Speech Recognition,” in Proceedings of International Conference of
Spoken Language Processing, Sydney, Australia, 1998, pp. 887–890.
[28] R.W. Picard, Affective Computing, MIT Press, Cambridge, Massachusetts, 2000.
[29] C. Chen and C. Mo, “A Method for Intelligent Fault Diagnosis of Rotating
Machinery,” Digital Signal Processing, Vol. 14, 2004, pp. 203–217.
[30] N. Oliver, E. Horvitz and A. Garg, “Layered Representations for Human Activity
Recognition,” in Proceedings of Fourth IEEE International Conference on
Multimodal Interfaces, Pittsburg, Pennsylvania, 2002, pp. 3–8.
[31] D.W. Albrecht, I. Zukerman and A.E. Nicholson, “Bayesian Models for Keyhole
Plan Recognition in an Adventure Game,” User Modeling and User-Adapted
Interaction, Vol. 8, No. 1–2, 1998, pp. 5–47.
[32] N. Blaylock and J. Allen, “Corpus-Based Statistical Goal Recognition,” in
Proceedings of Eighteenth International Conference on Artificial Intelligence,
Acapulco, Mexico, 2003, pp. 1303–1308.
Website
http://www.biometrics.org/
335
This page intentionally left blank
CHAPTER 11
Network Support: The Radio
Environment Map
Youping Zhao1, Bin Le2 and Jeffrey H. Reed1
1
Mobile and Portable Radio Research Group, 2 Center for Wireless
Telecommunications, Wireless@Virginia Tech, Bradley Department of Electrical
and Computer Engineering, Virginia Tech, Blacksburg, VA, USA
11.1 Introduction
The motivations for network support for cognitive radio are threefold:
1. With powerful network support, the requirements on cognitive radio user
equipment could be significantly relaxed because many computation-intensive
cognition functionalities can be realized at the network side. Distributed and
collaborative information processing over the network can reduce the workload
of single user’s equipment and speed up the adaptation process of cognitive
radio. This is an important strategy to facilitate the commercialization of cognitive radio technology, considering the many constraints imposed on costsensitive user equipment, such as limited battery power, signal processing
capability, and memory footprint.
2. Key cognitive functionality, such as incumbent primary user (PU) detection,
cannot be reliably accomplished by user equipment itself due to the shadowing
Note: Youping Zhao is supported through funding from the Electronics and Telecommunications
Research Institute (ETRI) and is advised by Jeffrey H. Reed. Bin Le is supported by the National
Science Foundation (NSF) under Grant No. CNS-0519959 and advised by Charles W. Bostian.
Any opinions, findings, and conclusions or recommendations expressed in this material are those
of the author(s) and do not necessarily reflect the views of ETRI or NSF.
337
Chapter 11
or fading effects of radio propagation and the practical system limitations of
the sensitivity, dynamic range, and the noise floor [1, 2]. The radio has to resort
to network support for many situations to solve the hidden node problem and
achieve the operational goals of cognitive radio.
3. Network support is critically important to the evolution of wireless communications from legacy radios to cognitive radios, and from the coexistence of various disparate radio networks to converged cooperative networks. As explained
in this chapter, network-enabled cognitive radio offers maximal flexibility to
the government, regulator, and service provider by supporting dynamic policies on spectrum access and utilization.
The radio environment map (REM) has been proposed as a vehicle of network
support for cognitive radio [3, 4]. The REM is an abstraction of real-world radio
scenarios; it characterizes the radio environment of cognitive radios in multiple
domains, such as geographical features, regulation, policy, radio equipment capability profile, and radio frequency (RF) emissions. The REM, which is essentially
an integrated spatiotemporal database, can be exploited to support cognitive functionality of the user equipment, such as situation awareness (SA), reasoning,
learning, and planning, even if the subscriber unit is relatively simple. The REM
can also be viewed as an extension to the available resource map (ARM), which is
proposed to be a real-time map of all radio activities in the network for cognitive
radio applications in unlicensed wide area networks (UWANs) [2, 5].
This chapter discusses the motivation and the important role of network support in cognitive radios, then introduces the REM and explains how the REM can
provide network support to cognitive radios for various applications. The main
concerns of network support and the related research issues are also addressed in
this chapter.
11.2 Internal and External Network Support
From the cognitive radio user’s point of view, the network support to the cognitive
radio can be classified into two categories: internal network support and external
network support. The internal network refers to the radio network with which the
cognitive radio is associated. Along with various communication services, the
internal network can provide some cognitive functionality as well. For example,
the cognitive radio network can provide location information and location-based
services to the user; it can also characterize the usage pattern of other users in the
neighborhood. The external network refers to any other networks that can provide
338
Network Support: The Radio Environment Map
meaningful knowledge to support the cognitive functionalities of the radio. For
example, a separate sensor network could be dedicated to gather information for
cognitive radio networks [6]. The external network could be a legacy network or
other cognitive radio networks.
Both internal and external networks can contribute to building up the REM and
can be employed in a collaborative way. For instance, location information needed
for a cognitive radio can be obtained either from internal network support through
a network-based positioning method for indoor scenarios, or from external network
support through the global positioning system (GPS) for outdoor scenarios.
As depicted in Figure 11.1, network support can be realized through a global
REM and local REMs. In this figure, the cognitive radio is symbolized as a
“brain”—empowered radio. The global REM maintained on the network keeps an
overview of the radio environment, and the local REMs stored at the user equipment only present more specific views to reduce the memory footprint and
communication overhead. The local REMs and global REM may exchange information in a timely manner so as to keep the information stored at different entities
current. In this figure, a regional REM can be aggregated from the combined
experiences of several local REMs.
Figure 11.1: REM. This provides
cognitive services to both the
associated internal networks and
a useful awareness of external
networks such as noncognitive
legacy systems.
Network
Support
for
Cognitive
Radio
External Network
Global
REM
Internal (Cognitive Radio)
Network
Local
REM
CR User
Equipment
Local
REM
Local
REM
Local
REM
11.3 Introduction to the REM
This section starts with an insightful analogy about cognitive radio, and then
introduces the REM as a cost-effective navigator for cognitive radios. Essentially,
339
Chapter 11
Location (x, y, z)
Geographical Information
Multipath Power Delay Profile
Radio Spectrum Profile
Radio Equipment Profile
Location-Specific Policy
Regulations
REM
Radio Environment Map
for Cognitive Radios
City Map for Travelers
Figure 11.2: City map versus REM. A REM provides services to local cognitive radios similar
to the way a city map aids a driver with local navigation. In contrast to the city map, the
REM database must be current to be useful.
the REM is a comprehensive spatiotemporal database and an abstraction of realworld radio scenarios [3, 4].
Similar to how a city map helps a traveler, the REM can help the cognitive
radio to know the radio environment by providing information on, for example,
spectral regulatory rules and user-defined policies to which the cognitive radio
should conform; spectrum opportunities; where the radio is now and where it is
heading; the appropriate channel model to use; the expected path loss and signalto-noise ratio (SNR); hidden nodes present in the neighborhood; usage patterns of
PUs1 and/or secondary users (SUs); and interference or jamming sources. Figure
11.2 shows such an insightful analogy. In fact, there are many commonalities
between transportation and telecommunication, such as the concept of throughput,
channel capacity, routing, signaling, rules, etiquette, and utilization efficiency of
1
PUs usually hold the license to use certain spectrum. Therefore, PUs have higher priority over
SUs. For more discussion on PUs and SUs, please refer to http://www.wireless-worldresearch.org/index.html.
340
Network Support: The Radio Environment Map
resources. Enlightened by this analogy, some lessons may be drawn from the field
of transportation and employed in the field of cognitive radio.
The REM can provide both current and historic radio environment information, so that most cognitive functionalities in the cognitive radio network can be
realized in a cost-efficient way. By leveraging the REM, cognitive radios can conduct spectrum sensing with prior knowledge rather than continually blindly scanning over the whole spectrum. Thus, observation time and energy consumption in
the radio front-end can be significantly reduced. By incorporating collaborative
information processing techniques, the costs of a cognitive radio system can be
further reduced by relaxing the requirements on transmission power, dynamic
range, and sensitivity of the individual radio device. By combining reasoning and
learning with data mining in a REM, network intelligence directly enables cognitive capabilities for its network nodes whether the subscriber radios are cognitive
or not. In this sense, even legacy networks can become cognitive by resorting to a
REM. The REM also supports a system-level solution to the fundamental cognitive radio technical challenges, such as SA, cross-layer optimization, the hidden
node and exposed node problem, load balancing across the network, opportunistic
spectrum access, and dynamic spectrum regulations and policy.
11.4 REM Infrastructure Support to Cognitive Radios
11.4.1 The Role of the REM in Cognitive Radio
The REM plays an important role in the cognition cycle of cognitive radio, as
illustrated in Figure 11.3. Both direct observations from the radio and knowledge
derived from network support can contribute to the global and/or local REM. The
radio’s environment awareness can be obtained from direct observation, such as
spectrum sensing, and/or from the REM. Reasoning and learning help the cognitive radio to identify the specific radio scenario, learn from past experience and
observations, and make decisions and plans to meet its goals. The global REM
and/or the local REM should be updated once action is taken or scheduled by the
radio to keep the REM’s information current.
The underlying techniques to support a REM include, but are not limited to,
database design, database management, database transactions (such as query,
search, and update), and data mining.
11.4.2 REM Design
The REM contains information at multiple layers, as illustrated in Figure 11.4. By
integrating various databases, the REM enables or supports cognitive functionality
341
Chapter 11
Reasoning, Learning
Awareness
Global and/or
Local REM
Decision,
Planning,
Adaptation
Network
Support
Observation
Figure 11.3: Role of REM for cognitive radio. The REM provides the infrastructure to
support the fusion of multiple cognitive services to the local subscribers.
for radios with different levels of intelligence. The REM helps a cognitive radio to
be aware of situations and make optimal adaptations according to its goals; for
legacy or hardware reconfigurable radios, the REM facilitates smart network
operations by providing cognitive strategies to the network radio resource management control. Just like the city map that is informative to every traveler, no
matter whether driving a car or taking the bus, the REM is transparent to the
specific radio access technology (RAT) to be employed regardless of whether the
subscriber radio is cognitive or not.
With the help of the REM, a radio can become cognitive of performance metrics, the application, topology, and network (routing), as well as the medium
access control (MAC) and physical (PHY) layers of communication stacks under
different and varying radio environments. For example, if the radio is used on the
battlefield, reliability and security are of high importance. Therefore, special
source coding, encryption, anti-jamming channel coding, frequency planning,
and routing algorithms could be employed accordingly. The REM can support
various network architectures: centralized, distributed, or heterogeneous networks,
or even point-to-point communications. It can also support collaborative information processing among multiple nodes for obtaining comprehensive awareness.
With the REM, a cognitive radio can choose an access network based on cost,
data rate, spectral efficiency, and many other performance metrics. The optimal
adaptation is subject to the constraints of various radio scenarios, such as
342
Network Support: The Radio Environment Map
Geographical
Information,
Terrain,
Building, etc
Service,
Regulation
WCDMA
WRAN
WLAN
LMDS
WCDMA
cdma
GSM
If cognitive radio is in downtown,
increases weight of spectral efficiency
If cognitive radio is at office,
prefers WLAN to WCDMA
Policy
Multilayer
Information
Integrated into
the REM
Radio
Equipment
Blind zone
Experience
Hidden node
Radio Environment MAP
Geographical
Information
Database
Radio
Tx/Rx
Database
FM/TV
Database
Cellular
Database
Radar
Database
WRAN
Database
Service
Database
Policy,
Regulation
Database
Memory of
Past
Experience
Figure 11.4: Integrating various databases for building up the REM.
available services, available spectrum, user policy, and the capability of the radio
equipment.
11.4.3 Enabling Techniques for Implementing REM
Figure 11.5 shows the interdisciplinary research nature of the REM design for
cognitive radios. To implement and exploit the REM, various technologies must
be employed, such as artificial intelligence (AI), detection and estimation, pattern
classification, cross-layer optimization, database management and data mining,
site-specific propagation prediction, and network-based ontology.
The next generation of cognitive radios and their required infrastructure are
enabled by a collection of many new technologies: data storage technology, very
large scale integration (VLSI), RF integrated circuits (RFICs), geographical information systems (GIS) (e.g., three-dimensional (3-D) digital map), terrain-based
343
Chapter 11
Figure 11.5:
Interdisciplinary research
for developing and
exploiting REM.
Database
Management
Data Mining
Artificial
Intelligence
Sensor
VLSI, RFIC
REM for
Cognitive Radio
Site-Specific
Propagation
Prediction
Cross-Layer
Adaptation and
Optimization
Networking
Positioning,
Estimation
Detection,
Classification
radio propagation modeling, and GPS receivers. For example, several electronics
manufacturers have recently developed handheld products with Gigabyte storage
capacity, and many cellular phones have already been equipped with GPS
receivers. Therefore, it is quite possible that future cognitive radios will have
enough memory to contain a comprehensive database such as the REM.
Radio propagation simulation tools can predict many important parameters
such as path loss, SNR or signal-to-interference and noise ratio (SINR), as long as
the geographical environment information is available [7]. Compared to the
empirical channel model-based prediction, the advanced site-specific radio propagation techniques and software tools using 3-D terrain maps have been successfully employed to provide much more reliable predictions for radio propagation
prediction and system planning and management [8]. This makes it possible to
embed several contours into the REM, such as the service contour, blind zone, and
interference region, which will enable cognitive radios to make decisions and
adaptations that overcome the most common user complaints.
Information to populate the REM can be obtained by:
1. Integrating and/or correlating various existing databases, such as GIS databases or radio equipment databases.
2. Sensing with collaboration among distributed nodes (which may or may not be
cognitive radios).
3. Observing from a dedicated sensor network and/or other external networks.
4. Probing the radio environment.
5. Estimating the radio propagation characteristics with software tools.
344
Network Support: The Radio Environment Map
11.5 Obtaining Awareness with the REM
This section explains the meaning and importance of SA to cognitive radios. It also
shows how radios become situation aware with the help of the REM. As discussed in
the previous section, the REM can present multidimensional information for cognitive radios, such as geographical environment, location and activities of radios, regulations, and policy of the user or service provider. One of the most important features
of the REM is that it is transparent to the specific application of cognitive radios.
11.5.1 Awareness: Prerequisite for Cognitive Radio
The REM presents comprehensive system-level knowledge for a cognitive engine
to exploit. The cognitive engine, which is the cognition core of the cognitive
radio, is typically implemented as a software system consisting of learning and
adaptation algorithms.
An interesting analogy between two intelligent agents—a taxi driver and a
cognitive engine—makes this clear (Figure 11.6). The comparisons of cognition
components between these two intelligent agents are listed in Tables 11.1 and 11.2.
The SA and performance measure for a taxi driver or pilot have been extensively
discussed [9, 10]. The capability of SA is one of the most important features that
differentiates a cognitive radio from an adaptive radio.
Taxi Driver’s capabilities
Cognitive Radio’s capabilities
Senses, and is aware of, its
operational radio environment
and its capabilities
Senses, and is aware of, the
driving environment and
capabilities
Can dynamically and
autonomously adjust its radio
operating parameters
accordingly
Learns from previous
experiences
Can dynamically and
autonomously adjust the driving
operation accordingly
Learns from previous
experiences
Deals with situations not
planned at the initial time of
learning to drive
Deals with situations not
planned at the initial time of
design
Figure 11.6: Capabilities of two intelligent agents: cognitive radio engine and taxi driver.
345
Table 11.1: Comparison of two intelligent agents: taxi driver and cognitive radio engine.
Agent type Environment
Performance measure
Sensors
Actuators
Taxi
driver [9]
Roads, other traffic,
pedestrians,
customers
Safe, fast, legal, comfortable
trip, maximize profits,
minimize collisions
Cameras, sonar,
speedometer, GPS,
odometer, accelerometer,
engine sensors, keyboard
Steering, accelerator,
brake, signal, horn,
display
Cognitive
radio
engine
Radio spectrum, other
traffic by PU and SU,
jammer, RF noise
and interference
Spectrum utilization, reliable,
fast, legal, cost-efficient,
low power consumption,
minimize interference
GPS, antenna, BER/
PER/FER, interference
temperature, quality of
service
Transmission power
control, MAC,
beamforming
346
BER: bit error rate; FER: frame error rate; PER: packet error rate.
Table 11.2: Three-level situation awareness.
Agent type
Level 1 awareness
Level 2 awareness
Level 3 awareness
Taxi driver
(or Pilot) [10]
Looks and perceives
basic information
Thinks about and understands
the meanings of that information
Uses the environmental understanding to anticipate what will
happen ahead in time and space
Cognitive
radio engine
Observes the RF spectrum,
and waveform and
surrounding activities,
with focused attention
Performs calculation, estimation,
reasoning, and understands the
radioactivity and what it represents
Performs prediction and planning
to anticipate radio performance,
network requirements, and user
needs
Network Support: The Radio Environment Map
It is apparent that both intelligent agents observe the environment and become
aware of their situations; make in situ decisions according to their observations,
anticipations, and experiences; and then execute intelligent adaptations to reach
their goals. They evolve by a spiral learning process (also known as the “cognition
cycle”) throughout their lifetimes.
Like the taxi driver, the cognitive radio may also have three levels of SA, as
shown in Table 11.2.
To appreciate the significance of three-level awareness, consider the following
scenario: suppose you are driving from your office to a restaurant for lunch. You
may make many “cognitive” decisions and adaptations according to your current
situation, observations, and previous experience:
●
●
●
●
●
●
If you have tight schedule for lunch, you may prefer to drive through a fastfood restaurant nearby.
If it is a sunny Friday, and you want to enjoy the lunch with your buddies, you
may carpool with your friends to a nice restaurant in another town.
If it is a rainy day, and the road is curvy, you may turn on the headlights and
drive slowly to avoid accidents.
As you drive, you still need to be aware of the traffic lights, road signs, and
speed limits.
You may look at a map for directions, or recall from your memory the most
convenient way (short-cut) to the restaurant.
When you learn from the radio news broadcast that there is an accident ahead,
you may anticipate that it will result in slow traffic and choose an alternate route.
This list could be extensive, considering the many other possible situations
and their adaptations for a simple automotive errand. Obviously, the cognitive
engine has both awareness and learning capability, but with an REM the ability to
predict radio performance is facilitated as readily as predicting the path distance
for the car trip.
11.5.2 Classification of Awareness
Awareness means the understanding of the situation [11]. Both geographical and RF
environment-related information (such as radio propagation characteristics, waveform, and spectral regulations) play major parts in cognitive radio knowledge [12]. In
addition, policy, goals, and contexts are also important issues for the cognitive radio.
347
Chapter 11
SA here means that the cognitive radio knows its current radio scenario, the
intent of the user, and the regulations with which it must comply. In summary, SA
may include, but is not limited to, the following key types of awareness:
(a) Location awareness: The cognitive radio knows where it is, in the form of latitude, longitude, and altitude, or relative location to some reference nodes.
(b) Geographical environment awareness: The cognitive radio knows the terrain
and geographical information related to the radio propagation and channel
characteristics. This awareness is critically important for a cognitive radio to
choose the appropriate spectrum, channel model, or RAT, antenna configuration, and networking techniques.
(c) RF environment and waveform awareness: The cognitive radio knows the
spectrum utilization, the existence of PUs and/or SUs, the topology of the
user group, the interference profile, and other RF characteristics that may be
of concern.
(d) Mobility and trajectory awareness: The cognitive radio knows its moving
speed and direction. For example, in conjunction with geographical awareness, the cognitive radio can know it is moving south along Main Street at a
speed of 45 miles per hour, and it can “foresee” the radio environment ahead,
such as the available channel after the user passes over the next hill or the
radio standards supported along the route.
(e) Power supply and energy efficiency awareness: The cognitive radio knows the
source of its power supply, the remaining battery life, and the energy efficiency of alternative adaptation schemes.
(f) Regulation awareness: The cognitive radio knows the spectrum allocation and
emission masks at specific locations and frequency bands, which are regulated
by government authorities such as the US Federal Communications
Commission (FCC).
(g) Policy awareness: The cognitive radio knows the policy defined by the user
and/or the service provider. For example, the user may prefer to use the wireless local area network (WLAN) from a specific service provider at some
locations for quality of service (QoS) or security reasons.
(h) Capability awareness: The cognitive radio knows its own capabilities as well
as those of its team members and/or the network. Such awareness may include
knowing which waveforms are supported, the maximum transmit power, and
the sensitivity of the cognitive radio.
348
Network Support: The Radio Environment Map
(i) Mission, context, and background awareness: The cognitive radio understands
the intent of the user, and knows what mode and volume of traffic it is going
to generate and what the impact of that traffic will be to the local networks.
The cognitive radio understands the QoS requirements, and how overhead
activities may trigger additional network traffic and latency.
(j) Priority awareness: The cognitive radio knows the user’s priorities and habits.
For example, the user may prefer to use low-cost services whenever possible
(e.g., to switch to WLAN from a third-generation (3G) system when entering
a Wi-Fi® zone), or may prefer reliability over cost.
(k) Language awareness: The cognitive radio knows the signs, ontologies, and
etiquette used among cognitive radios to communicate with each other.
(l) Past experience awareness: The cognitive radio remembers the past experience and learns from it.
Note that all of these items can be interrelated and employed together in various
ways. For example, the cognitive radio may adapt network topology, and make
dynamic spectrum access (DSA) decisions based on the user’s location, mission
goal, RF environment, regulations, and service priority. Table 11.3 summarizes
the various types of awareness a cognitive radio may have, the relative significance of each type of awareness, and how to obtain such awareness.
11.5.3 Obtaining SA
The cognitive radio can obtain SA through different approaches:
1. direct observation (e.g., through field measurement);
2. inference from network support (e.g., through a network-maintained REM); or
3. analysis of local terrain propagation models combined with existing database
structures defining known communication systems.
Using the previous analogy, suppose you are driving a car. You cannot just rely on
your own vision when you drive, especially under unfavorable weather conditions,
such as a snowstorm, or at night. The map complements your limited local vision
and helps you know the road conditions ahead and how far to the destination.
You can make an informed decision whether you need to stop at a gas station. You
may take extra caution because you know the road ahead will be winding. You
may schedule to have a dinner at the next rest area 10 miles ahead. In summary,
349
Table 11.3: Summary of situation awareness for cognitive radio.
Type
Significance
Approaches to obtain awareness
Location
High
●
●
●
●
350
Geographical
environment
Low to
medium
●
●
●
RF environment/ Medium
waveform
to high
●
●
●
●
Current status
GPS (or assisted GPS)
Network-based positioning
Landmark or RF fingerprint matching
Combination of inertial navigation and GPS
Many kinds of positioning techniques
are commercially available
Query and exploit GIS database
Terrain recognition
Site-specific propagation prediction
GIS database is available (e.g., from
US Geological Survey [13]). Many
site-specific propagation tools can
predict path loss, delay spread, and
service coverage
Radio transceiver database
Collective observations by cognitive radios
Sensor network
Field measurement
Microwave point-to-point radio, FM,
and TV station databases are available and
maintained by the FCC, which provides
radio station information, such as site
location and transmission power [14]
Mobility
and trajectory
Low to
medium
Estimate moving speed and trajectory of the
user by analyzing the change of locations over
a period of time and correlating with GIS
Can be addressed with current
technologies together with the REM
Power supply
and energy
efficiency
Battery:
high AC:
low
Measure the voltage and/or current of
power supply
Mature technique (e.g., the cellular
phone knows the source of its power
supply and the remaining power)
351
Policy
High
Defined by the service provider
and/or the user
Can be addressed with the policy
database in the REM
Regulation
High
Defined by the government authorities
Can be addressed with the regulation
database in the REM
Capability
High
Provided by the cognitive radios
and/or networks
Can be addressed with the capability
database in the REM
Mission/
Context/Background
environment
Low to
medium
●
Priority
Low to
medium
Defined by the service provider and/or the user
Can be addressed with the priority
database in the REM
Language
Medium
to high
Standardizing cognitive radio languages
and etiquettes
This is under development (e.g., by the
DARPA XG program [15])
Past experience
Medium
to high
Long- and short-term memory of experience
for recall
Can be realized with case-based decision
theory (CBDT) and other technologies
●
Using machine intelligence,
various sensors
Applying speech and/or image recognition
and understanding techniques
Common industry software tools
demonstrate some context-awareness,
and even interact with the user
occasionally
DARPA XG program: Defense Advanced Research Projects Agency NeXt Generation program; FM: frequency modulation.
Chapter 11
you easily obtain helpful context information. Therefore, when you drive, you can
examine the map, refresh your past experience if you had been there before, or
just take a “trial and error” learning approach if the map or previous experience
is insufficient.
For the cognitive radio, shadowing, fading, and Doppler shift are the most common degradation or distortions imparted by the channel. REM can provide the cognitive radio with channel characteristics associated with the location and direction of
a mobile user. Channel information can be obtained through observing instantaneous measurements of the environment as well as long-term measurements and
learning of the general characteristics of the environment. Once these channel measurements are available, models can be created to predict the performance of the
link. Of course, these models are stochastic and produce outputs that are random
variables. The variance of model outputs can be incorporated into the decision
process of cognitive radio. Furthermore, cognitive radio can take advantage of channel awareness for planning. With the awareness of shadowing and fading characteristics, cognitive radio may adopt the appropriate waveform (i.e., PHY and MAC
layer) to adapt to or take advantage of the propagation characteristics. For example,
in a multipath-intensive environment, cognitive radio may choose to apply multiple
input, multiple output (MIMO) techniques to improve its performance. Cognitive
radio can also anticipate emerging call drop due to multipath or shadowing and take
pre-emptive measures, such as switching on the backup power amplifier, increasing
the number of RAKE fingers, leveraging smart antenna resources or spreading gain,
altering the power-control policy, or making an intersystem handoff.
The REM can exist at the user’s terminal equipment (local REM) and/or at the
network level (global REM). The local REM may be unique to each user’s device.
Each cognitive radio may use its own local REM to memorize its past experience
as well as its current status. For example, the experience of a blind zone stored at
the REM of a high-quality radio could be different from that of a low-quality,
less-sophisticated radio.
As shown in Figure 11.7,2 by using a global REM, it is possible that even the
legacy radio network can be upgraded to support some cognitive functionality
and behave as if it were a cognitive radio. For instance, through a software
upgrade to the network-level radio resource management system, the legacy network can know the subscriber’s location, and the interference environment, and
2
In Figure 11.7, the term “radio” may refer to any type of radio device, even a cognitive RF identification (RFID) tag.
352
Network Support: The Radio Environment Map
Network
Support
for Cognitive
Radio
External Network
Global
REM
Cognitive Radio
Network
Legacy Radio
Less Powerful
Cognitive Radio
Local
REM
Cognitive Radio
Figure 11.7: Situation awareness through the REM. REM helps radios to become situation
aware by bootstrapping them with local accumulated experience, customized to each radio’s
functionality requirements.
instruct the radio to use the most effective PHY and MAC layers supported by the
user’s device. A simple radio with limited cognition capability can become more
capable by leveraging the REM-based network support. The local REM may
exchange information with the Global REM, for example, through some common
control channel.
11.6 Network Support Scenarios and Applications
This section illustrates the implementation of the REM for cognitive radio network support through the analyses of various application scenarios requiring different network structures and services [16]. It shows that REM-based network
support has the following fundamental features:
●
●
●
It fits into the core of cognitive radio functional architecture.
It is independent from specific network topology.
It is compatible with hybrid node technology and intelligence.
353
Chapter 11
11.6.1 Infrastructure-Based Network and Centralized Global REM
The centralized global REM can play an important role in many infrastructurebased radio systems, such as the IEEE 802.22 wireless regional area network
(WRAN) and cellular radio systems. For example, the 802.22 WRAN is the first
worldwide wireless standard based on cognitive radios. It is composed of WRAN
base stations (BS), repeaters, and consumer premise equipment (CPE). 802.22
systems are primarily targeted at rural and remote areas offering fixed wireless
access services. Figure 11.8 shows a typical WRAN scenario, in which TV stations,
TV receivers, wireless microphones, and public safety systems operating on
certain TV channels are PUs, and 802.22 system subscribers are SUs [17, 18].
Coexistence is the key requirement for 802.22 systems because the SUs must
avoid generating harmful interference to the PUs and/or other collocated SUs
[19, 20].
With a global REM maintained at a WRAN infrastructure, the WRAN BS can
know the location, antenna height, and transmit power of nearby TV stations; the
local terrain; the Grade B service contours of TV stations; the forbidden spectrum
Wireless
MIC
WRAN
Base Station
Public Safety
Radios
TV Station
WRAN Repeater
TV Station
WRAN
Base Station
Grade B
Contour of TV Station
TV Receiver
WRAN CPE
Figure 11.8: A typical radio environment for cognitive WRANs. It aggregates the knowledge
of all local wireless activity.
354
Network Support: The Radio Environment Map
used by public safety radios and satellite communications; the demographical distribution of TV receivers; and the available TV channels for use. In addition, it
knows the distribution and usage patterns of other WRAN service subscribers.
Such information helps the WRAN BS choose the best spectrum opportunity to
use at the optimal transmission power. Smart antenna techniques can be more efficiently employed at the BS and/or the CPE with the radio environment information from the REM.
Another example application of the global REM is cognitive WLAN supporting network-based interference management. Figure 11.9 shows the typical radio
environment of 802.11 WLAN in an office building, where various industrial, scientific, and medical (ISM) band interferences (such as microwave oven leakage,
cordless phone, Bluetooth devices, and ZigBee devices) co-exist [21]. A global
REM can be built up at the network infrastructure and used as a WLAN interference management tool. WLAN interference is detected, classified, and located by
the cognitive subscribers or access points (APs). Such information is stored or
updated in the global REM. In this way, the WLANs can cooperate to effectively
identify and/or report sources of interference, and make adaptations at the AP or
the subscriber device to avoid or mitigate the interference, such as changing the
channel, managing the diversity mechanisms, selecting proper MIMO schemes, or
switching off some APs. Note that no changes to the current WLAN PHY/MAC
standards are required; however, flexibility in the APs and/or subscriber devices
may be needed.
11.6.2 Ad hoc Mesh Networks and Distributed Local REMs
Local REMs can be used in ad hoc mesh networks that consist of cognitive radios.
The Wireless World Research Forum (WWRF) has established a work group
(WG3) to study cooperative and ad hoc networks as an integral and evolving part
of the future communication infrastructure, taking into account both wireless and
fixed aspects [22]. In such networks, the topology can be self-organized and optimized; various modulation and coding schemes can be adapted; smart antenna
and MIMO techniques can be selected dynamically; and information sharing and
collaboration methods, such as maximal throughput, improved reliability, or
extended transmit range, can be employed to achieve the goal(s) defined by the
network or participating radios. As pointed out in Chapter 7, for ad hoc cognitive
radio networks, distributed learning is another effective way of learning that can
increase the power of cognitive radios greatly. By sharing or exchanging the local
REMs, cognitive radios may learn to match the right routing protocol, either
355
356
Figure11.9: A typical REM of 2.4 GHz WLAN activity. This map provides interference avoidance and management services.
Network Support: The Radio Environment Map
proactively or reactively. Master nodes may be selected to be responsible for collecting the distributed local REMs and combining them into a complete REM for
the network. Such a complete map can be accessed by each individual node, in a
role similar to the routing table for an ad hoc network.
The Defense Advanced Research Projects Agency (DARPA) Adaptive
Cognition Enhanced Radio Teams (ACERT) program intends to create a “distributed radio” greater than the sum of its parts [23], as illustrated in Figure 11.10.
This scenario involves two teams of people forming two ad hoc networks. Both
intrateam and interteam communications may be needed. Under an ad hoc mesh
network scenario, the network support can be accomplished through distributed
local REMs, where the multipath profile and path loss experienced by each user
can be shared. Node and channel awareness can be realized by sharing and
exchanging local REMs.
11.7 Supporting Elements to the REM
The REM is essentially a database that stores information utilized in the decision
process of a cognitive network. As such, various elements in conjunction with the
REM are needed to create the cognitive network. These elements include the
learning, reasoning, and decision algorithms. Separating the REM from the learning, reasoning, and decision mechanisms helps to facilitate a system that can serve
a variety of radio types and networks and support a variety of applications.
Reasoning, data mining, and query capability to external databases can furnish
information for the REM; this composite capability is called the self-informing
system (SIS), and is shown in Figure 11.11. Decision-making outside of the SIS
is the key to integrating heterogeneous networks or even dissimilar radios within a
homogenous network. Given the information provided by the SIS, a radio with its
own decision-making capability can make decisions that are relevant to the radio
and/or network capabilities and the supported application. Even for non-cognitive
radios, a cognitive network can be achieved by interfacing the SIS with the network radio resource management system. Note that the SIS does not need to
reside at a central location. It can, in fact, be distributed among the nodes in the
network.
More complicated learning and reasoning techniques can be practically implemented on the network than on an individual cognitive radio. Data mining can be
used for extracting hidden information and predicting future events. For example,
data mining the Global REM can be used to find the usage pattern of PUs, SUs,
and other interferers from large quantities of data (i.e., observations over an
357
Network 1
358
Network 2
Figure 11.10: DARPA distributed radio scenario. Two ad hoc networks form around two teams of users with intrateam
and interteam command requirements, supported by a distributed REM database for both teams.
Network Support: The Radio Environment Map
Cognitive Network
Self-informing System
CR
CR
CR
REM REM
REM
Search
Query
Update
REM
Data Mining
External
Network
Intelligence
Network
Interface
Reasoning
Decision Making
Adaptive
Radio
Legacy
Radio
Figure 11.11: Cognitive radio self-informing system block diagram.
extended period of time), using such techniques as decision trees, neural networks, and neural-fuzzy systems.
As discussed in more detail in Chapter 12, the cognitive radio can employ
many learning methods, such as statistical learning, instance-based learning, reinforcement learning, and decision-tree induction learning. Two basic optimization
approaches may be employed by cognitive radios to optimize complex parameters: (1) the classical optimization based on the properties of the objective function; and (2) heuristic optimization. Common classical optimization routines
include the cyclic-coordinate method, the steepest descent method, and the quasiNewton and conjugate-gradient methods. Heuristic techniques include genetic
algorithm, simulated annealing, ant colony optimization, Tabu search, and neural
network, among others [4, 24]. Classical optimization is more analytically
tractable compared to the heuristic approach; however, neither of them can be
guaranteed to yield optimal solutions. As explained in Chapter 7, the cognitive
radio environment is typically characterized by multi-objective decision-making
(MODM). Without a single objective function measurement, it is difficult to apply
classic optimization theory to adapt the radio parameters. Genetic algorithms are
suitable for searching the optimal MAC/PHY/NET combinations and other
parameters for cognitive radios [4, 25]. Hidden Markov Models (HMMs) can be
359
Chapter 11
used for radio scene classification, case recognition, and for making meaningful
predictions based on past experience captured into the training data.
AI techniques are useful to fully exploit the REM to obtain SA and to enable
reasoning, decision-making, and self-learning. For example, a cognitive radio that
learns to foresee where and when a spectrum hole or interference will appear will
do better than one that does not. Just as evolution provides animals with enough
built-in reflexes so that they can survive long enough to learn for themselves, it
would be reasonable to provide cognitive radio with some initial knowledge as
well as an ability to learn. After sufficient experience in its environment, the
behavior of the cognitive radio can become effectively independent of its prior
knowledge. The incorporation of learning allows one to design a cognitive radio
that will succeed in a vast variety of environments.
11.8 Summary and Open Issues
This chapter has addressed the important role of network support in developing
cognitive radios for various application scenarios, including infrastructure-based as
well as infrastructure-less ad hoc networks. As a vehicle to providing network support to cognitive radios, the REM is proposed to be an integrated database that consists of multidomain information, such as geographical features, available services,
spectral regulations, locations and activities of radios, policies of the user and/or
service provider, and past experience. A REM can be exploited by a cognitive
engine to enhance or achieve most cognitive functionalities, such as SA, reasoning,
learning, planning, and decision support. Leveraging both internal and external
network support through global and local REMs presents a sensible approach to
implement cognitive radios in a reliable, flexible, and cost-effective way.
Network support can dramatically relax the requirements on a cognitive
radio device and improve the reliability of the whole cognitive radio network.
Considering the dynamic nature of spectral regulation and operation policy, the
REM-based cognitive radio is flexible and future-proof in the sense that it allows
regulators or service providers to modify or change their rules or policies simply
by updating the REMs accordingly. Because the REM is a comprehensive database
that contains multidomain information needed for cognitive radios, even a lowcost/low-complexity radio device can obtain basic cognitive functionalities by
referring to the REM.
As a system-level solution of the cognitive radio, the REM not only makes cognitive applications feasible, but, more importantly, it facilitates the evolution and
convergence of wireless communication networks with cost-efficient database
360
Network Support: The Radio Environment Map
development. The REM presents a smooth evolutionary path from the legacy radio
to the cognitive radio. The REM can also be viewed as a natural, but major evolution of radio resource management used in today’s commercial wireless networks.
The REM also has great potential in bridging or converging different wireless
communication networks, and thus facilitating the integration of various radio network architectures and access technologies, such as WRAN, Worldwide
Interoperability for Microwave Access (WiMAX), WLAN, wireless personal area
network (WPAN), Beyond 3G (B3G), and 4G systems, to support heterogenous
cognitive wireless communications [26].
Although the network-enabled approach for cognitive radio looks promising,
many technical issues currently remain open. Some important open questions include:
1. How can the global REM and/or the local REM keep its information current?
2. How current and at what level of granularity does the information contained in
the global and local REMs need to be in order to provide desired performance?
3. How much additional overhead is needed to keep the REM current, and what is
the constraint on network latency?
4. How well can the distributed local REMs serve an ad hoc mesh network?
5. How can the integrity, security, privacy, and reliability of the REM be assured?
Certainly there are many challenges ahead. Although these technical challenges
seem achievable, business, regulatory, and political challenges can be much more
difficult to predict and address.
References
[1]
[2]
[3]
[4]
[5]
J.H. Reed and C.W. Bostian, “Understanding the Issues in Software Defined
Cognitive Radio,” in Tutorial for 2005 1st IEEE International Symposium on New
Frontiers in Dynamic Spectrum Access Network, Baltimore, November 2005.
W. Krenik and A. Batra, “Cognitive Radio Techniques for Wide Area Networks,” in
42nd Design Automation Conference, Anaheim, CA, June 13–17, 2005.
Y. Zhao and J.H. Reed, “Radio Environment Map Design and Exploitation,” MPRG
Technical Report, Virginia Tech, Blacksburg, VA, December 2005.
J.H. Reed, C. Dietrich, J. Gaeddert, K. Kim, R. Menon, L. Morales and Y. Zhao,
“Development of a Cognitive Engine and Analysis of WRAN Cognitive Radio
Algorithms,” MPRG Technical Report, Virginia Tech, December 2005.
B. Krenik and C. Panasik, “The Potential for Unlicensed Wide Area Networks,” Wireless Advanced Architectures Group, Texas Instruments White Paper, October 2004.
361
Chapter 11
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
N. Sai Shankar, C. Corderio and K. Challapali, “Spectrum Agile Radios: Utilization
and Sensing Architectures,” in 2005 1st IEEE International Symposium on New
Frontiers in Dynamic Spectrum Access Network, Baltimore, November 2005.
T.S. Rappaport, Wireless Communications Principles and Practice, Prentice Hall,
Upper Saddle River, NJ, 2002.
G. Bauer, R. Bose and R. Jakoby, “Three-Dimensional Interference Investigations
for LMDS Networks Using an Urban Database,” IEEE Transactions on Antennas
and Propagation, Vol. 53, No. 8, Part 1, August 2005.
J.S. Russell and P. Norvig, Artificial Intelligence: A Modern Approach, 2nd Edition,
Pearson Education, Upper Saddle River, NJ, 2003.
http://www.2pass.co.uk/awareness.htm#SAdefinition
J. Mitola III and G.Q. Maguire Jr., “Cognitive Radio: Making Software Radios
More Personal,” IEEE Personal Communications, Vol. 6, No. 4, August 1999.
B. Fette, “The Promise and the Challenge of Cognitive Radio,” March 2004;
http://www.sdrforum.org/MTGS/mtg_40_sep04/fette_cognitive_radio.pdf
http://www.usgs.gov/
http://www.fcc.gov/mb/databases/cdbs/
http://www.darpa.mil/ato/programs/xg/
J. Polson, “Cognitive Radio Applications in Software Defined Radio,” Software
Defined Radio Technical Conference—SDR04, Phoenix, November 2004.
C. Corderio, K. Challapali, D. Birru and N.S. Shankar, “IEEE 802.22: The First
Worldwide Wireless Standard Based on Cognitive Radio,” in 2005 1st IEEE
International Symposium on New Frontiers in Dynamic Spectrum Access Network,
Baltimore, November 2005.
C.-J. Kim et al., “WRAN PHY/MAC Proposal for TDD/FDD,” Doc.: IEEE
802.22-05/0109r0, November 7, 2005.
Federal Communication Committee (FCC), “In the Matter of Facilitating
Opportunities for Flexible, Efficient, and Reliable Spectrum Use Employing
Cognitive Radio Technologies, Authorization and Use of Software Defined
Radios,” FCC NPRM 03-322, Washington, DC, December 30, 2003.
Federal Communication Committee (FCC), “Facilitating Opportunities for Flexible,
Efficient, and Reliable Spectrum Use Employing Cognitive Radio Technologies,”
FCC NAO 05-57, March 11, 2005.
Y. Zhao, B.G. Agee and J.H. Reed, “Simulation and Measurement of Microwave
Oven Leakage for 802.11 WLAN Interference Management,” IEEE International
Symposium on Microwave, Antenna, Propagation and EMC Technologies for
Wireless Communications, Beijing, China, August 8–12, 2005.
http://www.wireless-world-research.org/index.html
J.M. Smith, “Adaptive Cognition Enhanced Radio Teams (ACERT) Presentation,”
May 2005; http://www.darpa.mil/ipto/solicitations/open/05-37_PIP.htm
362
Network Support: The Radio Environment Map
[24] T. Wang, Global Optimization for Constrained Nonlinear Programming, Ph.D. Thesis,
Department of Computer Science, University of Illinois, Urbana, IL, December 2000.
[25] T.W. Rondeau, B. Le, C.J. Rieser and C.W. Bostian, “Cognitive Radios with Genetic
Algorithms: Intelligent Control of Software Defined Radios,” in Software Defined
Radio Technical Conference and Product Exposition, November 15–18, 2004.
[26] Wireless World Research Forum Working Group 6 (WWRF-WG6), “Cognitive Radio,
Spectrum and Radio Resource Management,” WWRF-WG6 White Paper, 2004.
363
This page intentionally left blank
CHAPTER 12
Cognitive Research: Knowledge
Representation and Learning
Vincent J. Kovarik Jr.
Harris Corporation, Melbourne, FL, USA
12.1 Introduction
Generally speaking, a cognitive radio is a radio system that can intelligently adapt
its behavior or operational characteristics in response to changes in the radio’s
internal state or external environment. This ability to adapt promises a range of
functional capabilities, including making the radio systems more “personal,” as
suggested by Mitola [1]. Some examples of internal states that may alter the operational behavior of the radio are:
●
●
Low Battery Power: A reduction in available power may result in the inability
to support multiple waveforms simultaneously. The radio may select and disable a low-priority waveform in order to conserve power or maintain operation
of a mission-critical waveform.
Component Failure: The radio may reroute around a failure point; it may provide an alternative service possibly at a lower data rate, or it may terminate a
lower priority service in favor of a mission-critical communication.
Some examples of external influences that can affect the operational behavior
include:
●
Co-Site Interference: Use of the same radio frequency (RF) by multiple radios
or a radio attempting to exercise multiple frequencies through a single set of RF
365
Chapter 12
equipment will diminish or negate successful communications. The radio may
select an alternate operating frequency or change other characteristics of the
waveform.
●
Background Noise: Normal RF background noise can impede the effectiveness
of communications on a given channel or frequency. Again, the radio may
sense the level of background noise and alter the frequency, initiate frequencyhopping, or take other adaptive transmission actions to enable end-to-end
communications.
Although, a cognitive radio system may exhibit intelligent behavior by adapting to changes in its environment, it is, nonetheless, following a fixed set of
behavioral guidelines. As long as the environmental and internal states remain
within the boundaries of its established guidelines, the radio system will continue to
successfully modify its operational behavior in response to those changes. However,
if the internal state or the environmental conditions fall outside the range of the
precoded patterns and responses, the radio system cannot identify a feasible solution
path to follow. Without a prescribed set of actions, the system typically ceases to
operate or vacillates between states in an attempt to find a stable operational mode.
Within a cognitive radio system are several common architectural
components:
1. Sensors or other methods for gathering information about the external environment and internal state of the system.
2. A corpus of knowledge that represents a set of behaviors to be performed in
response to some set or pattern of inputs and/or states.
3. A reasoning engine or algorithm that applies the knowledge to the current state
of the system and reaches one or more conclusions.
4. Some element that enables the system to change its operational characteristics
or, in some way, act upon the conclusions.
The reasoning engine modifies the operation of the system based on the application of knowledge to the combined state information. Reasoning is the process by
which the system has an existing set of knowledge, applies it to a current situation, and identifies a course of action.
Figure 12.1 illustrates the basic operational control of a cognitive radio. The
radio is performing some communications function represented by the operational
function box in the figure. The waveform implementation and general radio
366
Cognitive Research: Knowledge Representation and Learning
hardware information provide the internal state data of the radio system. One or more
sensors provide information regarding the environment in which the radio is operating. This state information provides input data to the reasoning engine. The reasoning engine applies the collection of knowledge to the current state of the system.
Based on the current combined state and the knowledge, the reasoning system
changes one or more operational parameters for the function being performed.
As illustrated in Figure 12.1, a cognitive radio can intelligently adapt its operational behavior in response to external and internal influences. However, such a
system has a critical-limiting factor: the system can adapt based solely on predefined behaviors as represented in the knowledge base. Thus, even though the radio
can react intelligently to external environment and internal state changes, it can
adapt its behavior only within the bounds of the previously defined knowledge—it
cannot adapt to new situations. In order to modify its behavior in response to new
situations, the radio must learn.
Learning entails not only the ability to sense and adapt based on precoded
algorithms or heuristics but also the ability to analyze sensory input, recognize
patterns, and modify internal behavioral specifications based on the resultant
analysis of the new situation.
External Environment
Tx
Rx
Sensors
External State
Internal State
Waveform
Configures
Knowledge
Parameters
Reasoning Engine
Operational Function
Sensors
Control
Radio System
Figure 12.1: Cognitive radio monitoring and control adds several additional functions to
the basic radio, labeled “operational function” in this figure.
367
Chapter 12
The basic machine-learning architecture is illustrated in Figure 12.2. The
architecture has a set of control parameters (e.g., rules, functions, algorithms, etc.)
that provide some control output to the external environment. The essential concept is that the machine-learning system has some perception of the environment,
gained through sensors, external data input, and other means, that provides the
baseline truth on which the system asserts some conclusion or action. Coupled
with the action is some predictive assertion regarding the anticipated impact or
change to the environment as a result of the action. The action, in turn, alters or
modifies the environment in some fashion. As the environment is modified, the
changes are received by the learning component and compared against the
expected changes. If the resultant changes are the same as the anticipated changes,
then, based on the proximity of the actual changes to the expected changes, the
system reinforces the parameters that led to the decision. If the resultant changes
are different than the predicted values, then the learning system will alter the
parameters of the decision process to more closely fit the actual results.
Figure 12.2: General
machine learning process
is built on a reasoning
system and a decisionmaking system.
Learning Component
Modified
Parameters
Predicted
State
Reasoning
Component
Decision
Parameters
Current State
Actions
External
System
Current State
So, how must the architecture shown in Figure 12.1 be modified in order to
support learning? Several changes must be made to the system, as illustrated in
Figure 12.3. In order to learn, there must be some method for assessing the operational performance of the reasoning engine and, based on that assessment, a method
for modifying the set of knowledge that is applied by the reasoning engine. So, as
illustrated in Figure 12.3, a learning algorithm is integrated into the radio system.
The learning algorithm, although potentially complex, has a simple mission: observe
the state of the radio system and the actions selected by the reasoning engine.
Compare the resultant state after the action selected by the reasoning engine is
performed. Then alter the knowledge base to reflect the success (or failure) of the
selected action. Alteration of the knowledge base can take multiple forms, from
modifying existing knowledge, to changing the action selection policy, to generating entirely new knowledge entries for the system to apply in subsequent actions.
368
Cognitive Research: Knowledge Representation and Learning
External Environment
Tx
Rx
Sensors
External State
Knowledge
Internal State
Waveform
Reasoning Engine
Parameters
Operational Function
Sensors
Control
Learning Algorithm
Radio System
Figure 12.3: Learning integrated into cognitive radio architecture.
The balance of this chapter is organized into three key sections and a summarization. Section 12.2 provides a brief overview of knowledge representation and
reasoning paradigms. Understanding the different representation mechanisms is
an important prerequisite to discussing learning algorithms because the algorithm
may be inextricably tied to the underlying representation and reasoning mechanism. Section 12.3 presents several approaches to machine learning and their
application within a cognitive radio system. Section 12.4 presents some of the
issues and trade-offs related to the implementation aspects of the learning systems. Finally, Section 12.5 summarizes the chapter.
This chapter provides a representative cross section of machine learning with
an emphasis on application within the radio communications domain. It does not,
however, provide an exhaustive overview of the multiple areas of machine learning, related research, and approaches; rather, some approaches are discussed in
depth whereas others may be discussed at a superficial level or not at all.
12.2 Knowledge Representation and Reasoning
A common theme throughout all of the machine-learning methods discussed in
this chapter—and, to a great extent, throughout all of machine-learning research—is
369
Chapter 12
the underlying tenet to enable a machine to react intelligently to some future situation based on knowledge it has gained through past experiences. Knowledge can
have multiple forms based on its type and the selected reasoning mechanism. How
the reasoning system represents and applies the knowledge underlying its decision
process has an impact on the learning process applied to improve the system. This
section looks at several common implementation methods and their potential uses
within a cognitive radio system.
In humans, knowledge is represented as a complex set of neural connections
that are activated in an associative manner. The set and pattern of the activation
represents a concept or memory, a contextual relationship, or an action or event.
Thus, when we are faced with a situation that we remember encountering before,
the set of neural connections activated by the current state also activates a set of
connections representing the action or response performed when last encountered.
The different sensory inputs interact to form a multidimensional perspective of
a concept or object in the physical world. For example, waking up in the morning
to the smell of bacon cooking can evoke a set of associated references. Based on
the smell of the food, we can visualize the bacon cooking in the frying pan and
remember the taste of the bacon we have eaten.
More significant than simply remembering the situation, however, is the ability to recall actions associated with a particular situation. Continuing with our
breakfast scenario, we may also remember the last time we cooked bacon and the
steps involved: getting the bacon out of the refrigerator, getting the frying pan
from the cabinet, getting tongs or another instrument to manipulate the bacon,
placing the bacon in the pan, flipping it over as it cooks, placing the cooked bacon
on a plate, and then placing the plate on the table. The key concept here is that, in
addition to remembering the objects or concepts, we also remember the process or
action sequence associated with achieving a particular goal.
Supporting a reasoning mechanism requires two critical capabilities. First, there
must be some form of capturing, categorizing, and storing the knowledge. Second,
the stored knowledge must be activated and applied to a given set of parameters or
conditions. Learning provides a method for extending the set of knowledge structures
without explicit programming on the part of a human developer. The types of learning that can be supported depend on the underlying structure of the knowledge.
Representing knowledge can take many forms, and in some cases, it may take multiple forms within a single system. For the objectives of this chapter, the knowledge
representation forms are limited to two basic categories: declarative and behavioral.
Declarative knowledge and reasoning refers to any method of representation
that asserts and reasons over descriptive factual knowledge about an object or
370
Cognitive Research: Knowledge Representation and Learning
entity. For example, the statement, “The car is red,” provides a simple declarative
fact—that the color of the car is red. In this example, the fact describes a property
or attribute about the car. This type of declarative knowledge is different than that
provided by the statement, “The car is owned by John.” In this statement, declarative knowledge is also provided, but, whereas the first example provides information that describes a property of the car, the second asserts a relationship between
two entities, the car and John. Thus, the second assertion of declarative knowledge
provides contextual or semantic knowledge about the car and its relationship to
other entities within the knowledge space. The key distinction is that the car’s
color is independent of contextual relationships and is a primitive property of the
car entity. The “owned” relationship, on the other hand, describes semantic connections between two objects.
Behavioral knowledge and reasoning is concerned with actions and the effect
they have on the knowledge system. The actions may be initiated by some agent
or actor within the system or they may be external events that alter the system’s
environment. In both cases, the event has some initial state or set of assertions at
the start of the action, some state at the end or completion of the action, and some
temporal extent over which the action occurs.
The remainder of this section briefly discusses several different representation
and reasoning approaches that intersect the above knowledge types. It should be
noted that a particular approach may encompass reasoning, representation, or
both. Furthermore, one reasoning approach may utilize different representation
approaches, depending on the situation.
12.2.1 Symbolic Representation
In a symbolic representation and reasoning system, extensible data structures capture the salient facts, descriptions, and properties associated with a concept. This
encapsulation of multiple descriptive information forms the symbol, or a unit of
knowledge. Symbolic knowledge representation can have any of multiple forms.
For example, knowledge may be represented symbolically as a conceptual entity
that has a set of properties describing the entity, or it may represent a semantic
association between two entities. For example, Figure 12.4 illustrates a simple
assertion in symbolic form expressing a transmits relationship between Radio-1
and Radio-2. This concise pictorial view captures the fact that Radio-1 transmits
some information or signal to Radio-2. The assertion may be represented in basic
data structures within a high-level programming language such as C, Java, or
list processing (LISP), to name a few.
371
Chapter 12
Each of the radios in Figure 12.4 can be represented by a data structure that
may have additional descriptive information or properties about each of the
radios. This is illustrated in Figure 12.5, in which Radio-1 now has additional
information associated with it that describes its operational frequency and the
power level of its transmitter amplifier.
Radio-1
TransmitsTo
Radio-2
Figure 12.4: Symbolic knowledge assertion: radio transmission represented as a from-to
“transmits to” graph.
Figure 12.5: Properties associated
with radio-1.
Radio-1
TransmitsTo
Radio-2
OperatingFrequency: 1900 MHz
Transmit Power: 1.2 W
The benefits of a symbolic knowledge representation approach include that it
is relatively easy for humans to understand and it is easily captured in graphical
and/or natural language notation. Symbolic knowledge representation also provides a more natural means of conveyance of the underlying information for
human understanding and interpretation.
Knowledge can be represented symbolically by using a variety of approaches,
including semantic nets, rules, frames, and objects, among others. The key underlying concept is that the representation formalism utilizes a symbolic form of
representing declarative knowledge.
As a cognitive radio system functions, part of the process, as described in the
prior section, is to observe the actual response or result of its interaction with the
environment, to analyze the result with respect to its expectations, and then to
adapt based on some learning algorithm. One of the applications of learning is to
extend existing knowledge based on newly observed data or patterns.
12.2.2 Ontologies and Frame Systems
Declarative knowledge is of little value when it is simply a disjoint collection of
facts and assertions. The data must be organized into a useful form in order to
support reasoning and learning. This organization is typically referred to as an
ontology. Ontology-based knowledge representation and reasoning for software
372
Cognitive Research: Knowledge Representation and Learning
radios have been investigated by a number of individuals, including Wang et al.
[2, 3], and in this text by Kokar et al. [4].
The classical is-a-kind-of relationship organizes classes or types of entities
into a type/subtype hierarchy similar to object-oriented programming languages.
The differences between the type and subtype can be structural or behavioral. For
example, a pager and a cell phone may be both classified as a “kind-of ” personal
communications device.
Ontology-based representation can be traced back to early frame systems [5],
and both have several aspects or viewpoints that can be applied to the organization
of a corpus of knowledge into an ontology. The Web Ontology Language (OWL)
by Bechhofer et al. [6] has been utilized in several efforts. For example, Baclawski
et al. [7] explore the use of ontology-based reasoning for communications protocol
interoperability.
Figure 12.6 illustrates a possible ontology of communications devices according to a type hierarchy. Two common types of devices, a cell phone and pager, are
described within this contextual organization.
The value of this type of organization from a learning system perspective is
that, as new concepts and situations are encountered, the existing knowledge
ontology can be searched to identify an appropriate classification for the new
CommunicationsDevice
OneWayCommunicationsDevice
TransmitOnlyDevice
ReceiveOnlyDevice
TwoWayCommunicationsDevice
HalfDuplex
FullDuplex
LocatioinBeacon
CellPhone
Pager
Figure 12.6: Communications device ontology.
373
PersonalCommunicationsDevice
Chapter 12
situation or concept. For example, if a Family Radio Service (FRS) walkie-talkie
is introduced into the system, it may be described as being used by an individual
capable of either sending or receiving at any given point in time. Using this
knowledge to guide the traversal of the knowledge hierarchy, the system would
follow the hierarchical links based on the known attributes of the new entity and
reach the conclusion that the cell phone radio system should be classified as a
PersonalCommunicationsDevice and a FullDuplex device.
Learning within an ontology-based system is performed by incorporating and
integrating new concepts into the existing set of entries. New concepts are incorporated into the ontology based on their similarity to existing entities; that is,
new information is processed by a classifier. The classifier analyzes the properties
and values associated with the new concept and compares it against the existing
set of concepts ontology within the ontology. Wherever there is a similarity, the
similar concept or concepts are candidates for associating the new concept into
the ontology.
This is similar to the method by which we incorporate new concepts within
the framework of our knowledge and experience. Using this approach, new entities can be linked into an ontology, enabling the system to extend its set of
knowledge. Note that, although the example shown describes physical entities,
the ontology can also represent related actions, events, or abstract entities such
as waveforms and the attributes associated with each of these types of entities.
Thus, organizing new knowledge can be performed across a range of conceptual
domains.
Also note, however, that learning based on ontology extensions is predicated
on the existence of an initial set of entities already organized into a hierarchical
network of concepts.
12.2.3 Behavioral Representation
Behavioral knowledge refers to the embodiment or description of actions that
an object or entity can perform or has performed. This may be behavior that is
executed in response to a simple stimulus, as in the case of object-oriented programming languages. This may be described within a neural net or genetic algorithm (GA) as set of patterns and actions, or it may be a more complex behavior
embodied within a set of rules or heuristics that allow additional options or
degrees of flexibility in the entity’s response to a given stimulus.
Although each type of behavioral knowledge may vary significantly in both
flexibility and the types of behavioral patterns it supports, each has the same
374
Cognitive Research: Knowledge Representation and Learning
fundamental objective. Each describes behavioral responses that an intelligent
system can exhibit in response to a stimulus that it receives.
Autonomously extending behavioral knowledge requires the system to incorporate a learning algorithm implementation that monitors the results of the system’s actions and adapts or modifies the decision parameters based on the specific
approach implemented by the learning algorithm. Typically, this is accomplished
by strengthening those decisions or actions that lead to achieving a system goal,
weakening those choices that did not lead to the attainment of the goal, or both.
12.2.4 Case-Based Reasoning
People faced with a new situation or problem typically attempt to find some prior
experience or situation that is similar and recall what actions were taken in
response to the situation. In some instances, there is a high degree of similarity
between the current situation and a prior experience in memory. When there is a
high degree of similarity, then the actions previously taken can usually be applied
to the current situation. In those instances where there is some similarity to prior
experience, then typically only some subset of the prior actions may be applicable,
or some may be applicable but require modification. This process of comparing
current situations to prior experience and then applying, potentially with adaptations, prior actions to the current situation is the fundamental approach of casebased reasoning (CBR). The origin of CBR is generally attributed to Schank [8].
Figure 12.7 illustrates the general architecture of a CBR system. Performing
CBR consists of several steps:
1. The current situation or state must be assessed and quantified. This entails collecting the set of external and internal sensory input, whether obtained autonomously or through external input, and coercing it into a canonical form that
is consistent with the set of previously stored cases.
Figure 12.7: CBR applies
prior experience to the
current situation.
Cases and Actions
Situation
Knowledge
Base
External
System
CBR
Actions
Retrieve and Update
375
Chapter 12
2. The current case is compared against the set of stored cases to identify previous experience that is similar to the current state.
3. The actions taken for one or more of the most similar cases are retrieved
and applied to the current situation. If there are no matching cases, then a
set of actions must be adapted from a similar case or new actions must be
proposed [9].
4. The case and the actions taken are updated or stored. If the case performed is
new, then it must be classified and inserted into the library of cases with the
appropriate semantic relationships to the other cases.
Implementations of CBR systems may vary as widely as the systems to which
they are applied. The underlying representation may be frame-based, organized
into ontologies, or based on symbolic knowledge representations. Regardless of
the underlying representation, the essential reasoning mechanism is to perform a
pattern-matching process that codifies the current state of the system, both internally and externally, and compares that against stored representations of previously encountered or codified states entered as a baseline set of knowledge.
This last point raises a key aspect of CBR. The accuracy and value of the reasoning process increases proportionately to the breadth and depth of the cases that
are available. In other words, to be effective, a case-based system must have a significant set of prior experiences from which it can search and identify similar situations that it can apply to the current state.
The second key aspect of an effective CBR system is its ability to adapt the
actions associated with a prior case that is close to but not a clear match for the
current situation. The ability to identify specific action steps requiring modification, to change those actions identified in a meaningful way, and to perform the
actions is a central CBR capability.
Depending on the underlying representation of the cases and the number of
cases stored in the knowledge base, the memory requirements for a comprehensive case-based system can be fairly significant. Another performance consideration involves the computational requirements to (1) perform the pattern matching
required to identify potential cases that are similar to the current state, and (2) perform the modifications necessary to the stored set of actions, when there is not a
highly similar match, in order to apply them to the current situation.
CBR would typically not be a candidate for low-end, resource-limited systems
due to its memory and processing requirements. It is more suitable for larger systems with adequate computing resources, such as fixed systems in vehicles, larger
aircraft, and ships.
376
Cognitive Research: Knowledge Representation and Learning
12.2.5 Rule-Based Systems
Rule-based (production) systems have a long history [10] and have been applied
to a variety of applications. A rule-based system has a knowledge base represented
as a collection of “rules” that are typically expressed as “if-then” clauses. The set
of rules forms the knowledge base that is applied to the current set of facts. Rulebased systems provide a method for representing inferential knowledge by using a
simple if-then form, which is relatively easy to state and understand. The rule paradigm is naturally understood by humans. The basic architecture of a production
rule system is shown in Figure 12.8.
As seen in Figure 12.8, the current state of the system is represented as a set of
facts or assertions in working memory. The corpus of knowledge is stored within
a set of production rules that form the knowledge base. The inference engine performs a pattern match of the antecedents or conditional portion of the production
rule against the set of assertions in the working memory. When there is a match,
the rule is tagged for possible execution. Because more than one rule may match
the current state in working memory, some mechanism is usually provided for
resolving the conflict and deciding which rule should be executed. The selected
rule is then executed, or “fired,” resulting in some action being performed or one
or more facts being asserted (added to) or retracted (removed from) the working
memory. Any external data that is represented in working memory is updated and
the cycle continues.
Figure 12.8: Basic architecture
of a rule-based system. The
production rule matches the
current state in working
memory to one or more rules
in the knowledge base.
Knowledge
Base
Rules
Working
Memory
Facts
Assert/Retract
Assert
377
Inference
Engine
Actions
External
System
Chapter 12
Figure 12.9 illustrates a simple power conservation rule that states that if the
radio has a waveform instantiated and the waveform is not actively transmitting,
then the system may conserve power by reducing the power level on the power
amplifier. The syntax shown is a common form for production systems. The
antecedents are expressed as a set of tuples in parentheses. The implied relationship between the tuples is usually AND. So, for the example rule in Figure 12.9,
there are two antecedent expressions, and both must be true for the rule to fire.
Figure 12.9: An example of rule-based reasoning.
If (?waveform is-instantiated)
(not (?waveform is transmitting))
=>
(set PA-power conserve)
The use of the question mark (?) in front of waveform denotes that the item is
a variable. So, any two-value tuple in the working memory for which the second
value equals “is-instantiated” is matched to the antecedent expression with the first
value of the tuple being bound to the ?waveform variable. Then, when the second
antecedent expression is evaluated, the ?waveform variable is set to the value
bound in the first expression and the working memory is searched for the pattern.
One of the shortcomings of the rule-based paradigm, however, is that it typically has no means to introspect the knowledge within the system. In other words,
the system’s knowledge provides guidance regarding the domain of the system
and responds based on the set of knowledge and the current state of the environment. However, the rule engine cannot scan through the rules to adjust them, add
rules, or delete rules. To accomplish these activities, an additional set of reasoning
must be implemented. Thus the learning algorithm would be applied to monitor
which rules were applied in a particular situation, assess the success of the decision process based on the actual outcome versus the predicted outcome, and then
have at its disposal access to the rules in a form that the learning algorithm can
understand and process.
12.2.6 Temporal Knowledge
Temporal reasoning provides the ability for a system to reason about its operational characteristics within the context of time. More than simply a discrete set of
points, temporal reasoning that captures relationships between temporal intervals
provides a basis for qualitative temporal reasoning that does not require assignment of specific time values. The notion of temporal interval logic proposed by
378
Cognitive Research: Knowledge Representation and Learning
Allen [11] comprises a set of 13 relationships that capture the relationship
between any two temporal intervals. For example, interval A may overlap interval
B, implying that A starts before B and A ends sometime between the start and end
of B. A number of researchers have developed additional aspects of temporal
interval logic, including Kovarik and Gonzalez [12], resulting in a rich variety of
temporal representation and reasoning approaches, as described by Allen [13].
Dynamic spectrum utilization can be mapped into a temporal representation
paradigm by mapping the times associated with observed frequency usage. Once a
map of spectrum usage over time has been built, intervals of time can be identified
when a particular portion of the spectrum is underutilized. The temporal “holes”
can then be used to more intelligently allocate spectrum in a collaborative fashion
across a number of devices. This concept is illustrated in Figure 12.10.
Temporal reasoning also comes into play when considering aspects such as
co-site interference and concurrent requests for physical and processing resources.
3) Assign Slot to Requestor
4) Assignment Is a Temporal
Constraint Satisfaction
Problem
Frequency
1) Spectrum Request
Co-Site Interference Constraints
2) Search for Available Time Slot
in Frequency Requested
Resource Constraints
Time
Figure 12.10: Dynamic spectrum assignment based on temporal mapping of frequency usage.
12.2.7 Knowledge Representation Summary
Section 12.2 has thus far presented a brief overview of representation and reasoning paradigms. Each of the different implementation approaches has benefits and
379
Chapter 12
limitations. Although there are multiple reasoning paradigms from which to
choose, several key tenets can be asserted regarding the use of a reasoning system
within a cognitive radio system.
First, even though any one of the approaches described, as well as those not
covered in this section, will provide a measurable degree of intelligent behavior
within a cognitive radio, no single representation and reasoning system is capable
of embodying all of the types of knowledge and reasoning expressiveness required
for the range of operational situations that a radio system must be capable of
handling.
Second, although each reasoning mechanism provides a degree of adaptability
to operational situations not explicitly defined, when any of the mechanisms
encounters a totally new situation or scenario, the ability of the system to reason
and decide on an appropriate behavior or response drops off sharply. This brittleness at the edge conditions leads to the pattern of oscillation between stable and
unstable states or termination of operation because the system has encountered a
situation for which it does not have any precoded knowledge.
Finally, the performance of each of the reasoning systems discussed is highly
dependent on the breadth and depth of the knowledge codified within the system.
The more knowledge that is encoded within a system, the better it performs. This
axiom has its consequences, of course, in that the cost in terms of memory and
computational resources required to support a specific reasoning system may be
significantly more than is available on some resource-limited systems.
The next section discusses different learning mechanisms. Each of the
methods can typically be applied to different representation and reasoning implementations, as discussed in this section.
12.3 Machine Learning
Research in machine learning has grown dramatically in the past decade, with a
significant amount of progress along several research paths. Multiple strategies
exist for implementing machine learning, each with some degree of benefits as
well as drawbacks both in general and in the context of a cognitive radio system.
This section addresses common learning approaches and provides some analysis
of their applicability to cognitive radio systems.
One of the important aspects of the learning mechanism is whether the learning performed is supervised or unsupervised. In a supervised learning system, as
the name implies, learning is performed through a set of predetermined conditions, and the system is trained to select the right choice or outcome. In effect, the
380
Cognitive Research: Knowledge Representation and Learning
learning mechanism learns to associate a particular set of input stimuli with a
learned response or output. In the context of a cognitive radio system, either technique may be applied as the radio’s method of learning. Supervised learning may
take the form of a dialog between the operator and the radio, in which the radio
may develop some new assertion or operational behavior model based on the
learning mechanism within the radio and then ask for confirmation from the
operator that its conclusion is correct.
Conversely, the radio system may extend its knowledge through the
learning algorithm and simply add the new knowledge to its existing base of
knowledge assertions and behaviors. In operational radio systems, there will likely
be some level of protection or guard to limit the degree of behavioral modifications that a cognitive radio may incorporate without external verification and
validation.
The potential problem with autonomous or unsupervised learning is that the
learning process may result in many wrong choices before a feasible decision path
is developed. This can be computationally intensive and time consuming.
Learning algorithms, such as reinforcement and temporal difference, which are
discussed in Sections 12.3.5 and 12.3.6, respectively, are applicable to this type of
learning.
The following sections provide an overview of several types of learning mechanisms. As previously noted, the focus is on those learning mechanisms that are
applicable to cognitive radio systems. Consequently, not all learning methods are
covered within this chapter.
12.3.1 Memorization
One of the most basic learning mechanisms is memorization, or rote learning. This
approach captures a sequence of steps or a response to a specific set of conditions
and then, when the same task or set of conditions is again encountered, the memorized responses are applied. This can be an effective method of learning if the range
of situations encountered by the radio system is limited and well defined. Incomplete
and partial data and situations, however, are not amenable to rote learning.
Rote learning forms a set of steps that represent the observed or described
sequence of actions for a given set of conditions. Applying the sequence learned
requires the system to take a set of measurements, sensor data, or descriptive input
data provided by an external source and then use the set of information as an associative key to look up the information within an internal database of patterns.
When the matching set of descriptive data is found, then the set of associated
actions are activated and applied to the current state.
381
Chapter 12
An example of memorization would be the selection of a particular waveform
or specific adjustments to the operating parameters of a waveform in response to
measured interference. Although the memorization would allow for multiple
response sequences, each associated with a specific interface value, each sequence
of actions would be explicitly tied to the measured values. Memorization does not
allow for generalization of responses based on similar responses to different
measured values. This capability could be thought of as a computational implementation of conditioned response, as pioneered by Ivan Pavlov.1 When the radio
senses some specific combination of stimuli, it performs some specific conditioned response based on the association of the response to the stimuli over time.
The problem is, of course, that if the set of stimuli that may be sensed by the radio
are insufficient to delineate two distinct conditions that may exist given the same
set of stimuli, the radio will execute a learned behavior that may not be correct for
the actual condition.
Due to the explicit and inflexible nature of memorization learning, it is best
suited for relatively fixed sequences of actions in which there is not a great deal of
variation in the set of choices. Examples would be the ordered set of steps associated with radio start-up and initialization.
12.3.2 Classifiers
Classifiers refer to algorithms that analyze two or more units of data or knowledge
and identify similarities or patterns in the structure and content of the data.
Classifiers are applied to areas such as data mining and, in the case of declarative
knowledge structures, provide a learning mechanism by extending an ontology
with new concepts based on the similarity of the new concepts to those already
within the system.
Similarities can be discovered between two distinct pieces of declarative
knowledge by a number of methods. A classifier may compare the set of properties associated with each of the concepts and, based on the similarities of the two
concepts, item B may be deemed to be subtype of item A. This would result in the
insertion of item B into the ontology as a subtype or subclass of item A.
The process of quantitatively assessing the similarity between two concepts can
be attributed to Tversky in his “Features of Similarity” paper [15]. A computational
1
Pavlov’s research on conditional reflexes greatly influenced not only science, but also popular culture. The phrase “Pavlov’s dog” is often used to describe someone who merely reacts to a situation
rather than using critical thinking [14].
382
Cognitive Research: Knowledge Representation and Learning
implementation of a common method for assessing the similarity between two
objects is to take the set of properties associated with each of the objects and create an n-dimensional space, with each dimension representing a property, and then
assign a number between 0 and 1 representing the value of each property for each
object.
Once this structure has been built for each object, the center-of-gravity point
can be calculated as the sum of the distances between each of the property values
in the n-dimensional space. Then, the relative similarity between the two concepts
is the distance (i.e., difference in magnitude) between the two center-of-gravity
points. The shorter the distance, the more similar the concepts.
In addition to classifiers being applied to declarative knowledge structures in
an ontology, they can also be applied to identifying similarities and patterns in
declarative structures representing temporal relationships. Learning as applied to
temporal knowledge enables the radio system to enhance or extend its repertoire
of known patterns of behavior and then apply those patterns to new situations. For
example, the system may observe a particular pattern of interference on a regular
set of frequencies at a particular time of day. Or, a system may observe that a particular pattern of activities and events precedes the initiation of a communications
activity that is of interest from a signal analysis perspective. Repeated observation
of the temporal patterns reinforces the validity of the set of observed temporal
events. The stored temporal pattern can then be applied by matching observed
events against the collection of known temporal patterns to predict a specific
activity and initiate an appropriate action or countermeasure based on the temporal prediction.
12.3.3 Bayesian Logic
Statistical learning methods employ some method of probability of a given outcome for a given set of input stimuli. The system matches a set of active input
stimuli to one or more sets of statistical functions having the same input parameters, and then applies the function to the input values, thus generating an expected
outcome, course of action, or classification assignment. The probabilities applied
to the calculation affect the result.
Learning is accomplished through the system observing the actual outcome as
opposed to the expected outcome and adjusting the weights accordingly. Here the
range of statistical functions that can be applied is significantly large.
One of the fundamental behaviors of a reasoning system is the tendency to
alter the prediction or expectation of the outcome of an event based on prior
383
Chapter 12
history. This concept is the fundamental idea of Bayesian logic. The probability
behind Bayes’ theorem is shown in Eq. (12.1):
P(f | x ) P( x | f) P(f)
P( x | f) P(f) P( x |~f) P (~f)
(12.1)
The event or occurrence of interest is represented as , and x is the observed
phenomenon or evidence of the occurrence. P(| x) is the probability that the
event has occurred given the observation x or, in other words, a measure of the
reliability of the assertion that has actually occurred given the observation x.
The probability of the observation for a specified set of the occurrence of interest
is P(x | ), and P() is the overall probability of the occurrence within a sample
population. Conversely, P(x|) is the probability that the observation
will manifest itself in a sample population that does not contain the event or
occurrence of interest, and P() represents the percentage of the set that does
not contain the event or occurrence of interest. An example in the context
of a cognitive radio illustrates how Bayes’ theorem might be applied as a learning
mechanism.
Suppose the problem domain is predicting whether there will be interference,
, encountered for a given geographical region. Further, assume that the radio has
one or more sensors that can detect the interference. When the interference is
detected, a memory location is set to a specified value, flagging the detection that
provides the evidence, x, of the interference phenomenon.
As a starting point, assign a value, P(), representing the probability that
there actually is interference in a given region. The probability that the interference is not present would thus be P(). However, because detection equipment
is not perfectly accurate all the time, there is a finite probability that it will detect
interference when it is present, P(x|), and also a finite probability that it will
detect interference when it is not present, P(x|), which is also called a false
positive.
Assume that the detection equipment is calibrated in the lab and is shown to
be accurate 80 percent of the time. So, for every 100 times that interference is
present, 80 will be detected and 20 will not. Assume also, however, that for every
100 readings when there is no interference, the detector registers interference
(false positive) 10 percent of the time. Now assume a total population of 10,000
samples of the environment and that the probability of interference is 30 percent.
So, of the 10,000 samples, 3000 are likely to be samples containing interference
and 7000 are not.
384
Cognitive Research: Knowledge Representation and Learning
Inserting these values into Eq. (12.1) yields Eq. (12.2):
P(f / x ) 0.8 0.3
0.8 0.3 0.1 0.7
0.774
(12.2)
Therefore, based on an observed interference phenomenon, it can be asserted with
roughly a 77 percent degree of confidence that the interference is actually present.
Bayesian logic can be adapted as part of the learning mechanism through
adjustment of the weights based on observed phenomenon. The limiting factor in
general use is the required verification of the prediction in order to adjust the
probabilities. Thus, the learning mechanism must have either external input or
some internal method of verification of the predicted value in order to adjust the
weighted values. Nonetheless, Bayesian logic can provide valuable reasoning support within a cognitive radio system by tempering the predicted outcome based on
prior experience.
12.3.4 Decision Trees
A decision tree, as the name implies, is a directed graph consisting of a hierarchical set of nodes connected via arcs, where each node represents a choice or decision, and the arcs leading from that node to the next decision node represent the
set of possible choices for a given node. A decision tree can be visualized as a
sequence of choices in which the path taken through the choices from the starting
point to an endpoint is governed by the choices made at the starting node and each
successive node. The root node provides the starting point and the leaf nodes contain the decisions or actions to be taken.
Decision trees may be simplistic in nature, representing a fixed procedural
process, such as problem diagnosis, or may be more complex, applying Bayesian
reasoning to the decision process governing the transition between the nodes of
the graph.
This type of reasoning approach is referred to as a solution space search problem. That is, the system generates a set of possible alternative actions for a given
state. Each of the generated alternatives is then explored, alternative actions are
generated for each of those states, and so on. As each alternative is generated, its
weight is calculated, subsequent alternatives are generated, and a total weight or
probability is calculated. Thus, for any given node in a decision tree, the total
weight for the path leading to that point is maintained.
385
Chapter 12
The calculation of the weights can be dynamically adjusted based on an
observed state compared to an anticipated state. Thus, the system can learn to
respond differently to different scenarios and states. The learning aspect can be
extended further by enabling the system to generate wholly new actions for a
given node. This enables the radio system to not only learn new behavioral
responses based on performance observations, but also to extend the range of
choices.
12.3.5 Reinforcement-Based Learning
Reinforcement-based learning provides a method for assessing the success of a
particular action, where success is defined as the actual outcome of the event
being the same as (or near to) the anticipated or desired outcome. Based on the
degree of success, a reward weight is assigned to the action, thereby reinforcing
the selection of that particular action if the system finds itself in the same state at
some future time.
There have been a number of advances in reinforcement learning [16].
Reinforcement learning is represented by a series of states as a directed graph.
As the system moves through the sequence of states by applying some policy for
selecting a path or alternative, the degree of success attained in achieving the system’s goals, both short term and long term, are measured. Based on the degree of
success, the policy is reinforced.
At each point in time, t, the system is in some state, s. The system has a set of
actions, A, from which it selects a specific action, a, where a A to perform in
the given situation with the objective of transitioning to some goal state, s, at the
completion of the action. In order to select an action, the system has a set of policies or rules for deciding which action to apply. The selection of a specific action,
a, given a state, s, by a policy, , can be represented as a (s).
Based on the degree of success, the system assigns a reward to the action performed, which, in effect, is a reinforcement of the policy selection rules. The
reward function, R(s,a,s), provides immediate feedback to the selection process.
So, as the system selects actions that achieve the goal state, the policy rules that
contributed to that selection process are reinforced and the probability of selecting
the same action again, if the same set of conditions is present, is increased.
Each correct selection of an action provides an immediate reward. Over the
course of several actions, a cumulative reward value is calculated. This provides
the mechanism for strengthening the selection of a sequence of actions. In
essence, it is reinforcing the ordered set of actions for the attainment of a final
objective or goal. Therefore, two methods influence the selection of an action.
386
Cognitive Research: Knowledge Representation and Learning
One is the selection of the current action based on the reinforcement of that action
successfully achieving the desired objectives for the current situation, and the
other is the influence of the successful attainment of the end objective providing
reinforcement of the action as a part of the ordered set of actions.
This attainment of a long-term goal is typically represented within the reinforcement paradigm as a cumulative value that is the sum of each of the immediate reward values. Each of the immediate reward values that are part of the sum is
discounted by applying a factor, t, to the reward value. The cumulative reward is
then expressed as shown in Eq. (12.3).
CumulativeReward ∑ gt R( s, a, s)
t 0
(12.3)
where 0 t 1.
The effect of the discount factor is to balance the reinforcement value between
realization of a short-term goal (i.e., whether the selected action yielded the
expected state, given the current state) versus the long-term goal (i.e., whether the
selected action contributed positively to the attainment of the long-term goal).
Thus, for any given state and action pair, there is a computed value function
for the policy based on the current state, denoted as V(s). Given that there may be
multiple states, s, that may be the result of performing the selected action, there
is a probability assignment that is made for each of the possible states. The probability of ending in state s, given a current state of s and that action (s) is taken,
is represented as P(s| s, (s)). This probability is then combined with the sum of
the short-term reward associated with the current action, R(s, (s), s), and the
cumulative discounted reward, V (s). This computed value for each possible
path is summed together, as shown in Eq. (12.4):
V p ( s ) ∑ s′ P( s | s, p( s )) ⋅ [ R( s, p( s ), s) gV p ( s)]
(12.4)
The overall objective of reinforcement learning is to compute the optimal policy, *, for each iteration of the process that maximizes the total reward shown in
Eq. (12.4).
Note that the term policy rule can be realized within the reinforcementlearning system in a wide variety of ways. The policy may start out as nothing
more than a random selection process which, over time, evolves to select the
appropriate action based solely on matching the pattern of current state variables
387
Chapter 12
to those patterns that were previously encountered, stored, and reinforced because
the system chose the right action.
Figure 12.11 illustrates the application of this approach. The radio system is in
some current state, s, and a policy, a, has been selected. There are four possible
states that may be entered upon the completion of the action. For each of these
possible states, a probability, P(s| s, (s)), has been assigned. The immediate
reward, R(s | s, (s), s), for each of the subsequent states is shown, along with the
estimated value of each of the states, V (s).
Figure 12.11:
Reinforcement learning
applied to policy
selection.
R(s, π(s), s)
V π(s)
2
0
S2’
0
4
S3’
1
10
3
6
P(s/s, π(s))
S1’
Policy
Selection
0.1
0.2
S
a
0.6
0.1
S4’
Assume that the discount factor, , is equal to 0.6. This would yield a new
estimated value of V (s) as:
V p (s) 0.1(2 0.6(0)) 0.2(0 0.6(4)) 0.6(1 0.6((10)) 0.1(3 0.6(6))
p
V (s) 5.5
(12.5)
This value becomes the new value of the selected policy for the current node
V (s). This process of modifying the policy value of the current node based on the
probable path and subsequent nodes is referred to as backing up the policy value.
Problems arise with reinforcement learning because, in order to function, a
complete model of the system is required that provides the transition probability
between states and the reward function. Defining the model to the level of detail
required may not be feasible for a cognitive radio system.
388
Cognitive Research: Knowledge Representation and Learning
Finally, in addition to the infeasibility of specifying the complete model
description, there is the issue of the algorithm performance. Performing the policy
iteration calculations requires O(n3) time where n is the number of states in the
system.
12.3.6 Temporal Difference
Temporal difference also applies a reinforcement algorithm to the policy selected
based on the degree of success. However, the temporal difference algorithm does
not require a priori model of the sequence of possible states, as the standard reinforcement algorithm does. Thus the temporal difference algorithm builds the state
representation “on the fly” and does not require the back-propagation used in the
reinforcement algorithm to update policy weights.
Introduced by Sutton [17], the general temporal difference algorithm is
referred to as TD(). There is a variant of temporal difference learning, TD(0),
that performs the computation of simple backups as the system transitions through
the states. As with reinforcement learning, as described in Section 12.3.5, the system is in state s and, based on the selection of an action based on a policy associated for state s, a (s), performs a probabilistic transition to a new state, s.
The transition produces the immediate reward R(s, a, s). The TD(0) algorithm
updates the value function by using Eq. (12.6):
V p( s ) : (1 a)V p( s ) a[ R( s, a, s) gV p( s)]
(12.6)
In Eq. (12.6), is a learning rate parameter. The initial value of this parameter is usually between 0.01 and 0.5 and it must, over time, reduce down to zero to
allow the TD(0) algorithm to converge. As can be seen in the equation, as the
learning parameter reaches zero, the latter portion of the function that incorporates
the immediate reward and the discounted value of the next state is algebraically
eliminated, and the system converges to the state value V (s).
Thus, as a node is visited multiple times, the effect is the same as performing a
simple backup, as with the reinforcement-learning algorithm in Section 12.3.5.
The benefit of the temporal difference algorithm is that the policy value can be
calculated without an explicit model. The set of states traversed as a result of the
actions and the environment become the model. This can be accomplished
because the empirical data gathered as the system traverses the set of states
provides the basic data required to compute a probability function for each node
traversed.
389
Chapter 12
A second advancement of the temporal difference algorithm introduced by
Sutton [17] over the reinforcement algorithm is that, rather than maintaining a
separate value, V (s), for each state s, the value function can be stored as a
neural network or some other method for providing a differential approximation.
Thus, the value function can be represented as V(s, W), where W is a vector of
weights that are modified based on observed state versus expected state. This prevents the ability to directly assign a value to a state. However, the weights specified in W may be adjusted such that V(s, W) fits closely to the desired value
through the use of an error function. The temporal difference error function is
shown in Eq. (12.7):
J (W ) 1
2
(V p ( s, W ) [ R( s, a, s) gV p ( s, W )])2
(12.7)
This equation captures the temporal difference error as one-half the
squared difference between the current estimated value of state s and the backedup value. Thus, the objective is to modify the set of weights, W, such that the
temporal difference error, J(W ), is reduced. By differentiating the equation
with respect to W and limiting only the first occurrence as being adjustable,
the general weight assignment function is obtained. This is shown in
Eq. (12.8):
W W awV ( s, W )(V p ( s, W ) [ R( s, a, s) gV p ( s, W )]))
(12.8)
The term represents the gradient of V with respect to the weights W. The
term represents the step size to be taken to reduce the error term calculated as
the temporal difference error.
12.3.7 Neural Networks
Neural networks and GAs also rely on the reinforcement of a decision or selection
based on the actual result or outcome of a decision. (See Section 12.3.8 and
Chapter 7 for a discussion on GAs.) The neural net typically has a vector of input
values and a vector of output values. An intermediate layer couples the input and
output values together, propagating the input values along a set of connections
from the input to the output.
Essentially, each node in the neural network takes an input vector, A, and
applies a weight vector, W, to perform the propagation of the input value along the
link to the next layer in the network. This propagation is typically tempered by a
390
Cognitive Research: Knowledge Representation and Learning
constant or bias value, b. Thus, for an individual node, the propagation output, A,
of a given node, j, can be expressed as shown in Eq. (12.9):
⎤
⎡n
A j ⎢ ∑ ai wi b j ⎥
⎥
⎢i 0
⎦
⎣
(12.9)
When the output value, Aj, is greater than some threshold value, Tj, the value
is propagated along the output, as illustrated in Figure 12.12. Learning, within the
context of a cognitive radio, involves the adjustment of the threshold value, the
bias value, or the weight associated with a node.
Figure 12.12: Neural network node
illustration. Shown are the input vector A
(with elements a0 … an); weight matrix W
(with elements w0j … wnj); bias value, b;
and output.
bj
a0w0j
a1w1j
.
.
.
Aj > Ti
Aj
a(n-1)w(n-1)j
anwnj
A neural network is formed when a collection of individual nodes is organized
together in a multilayer fashion, as illustrated in Figure 12.13. The neural network
has a set of input points. The values sensed or input through these points is propagated forward, through the middle layer (also referred to as the hidden layer), to a
set of output points. The output points activated by the forward-propagation are
then compared with the actual value (i.e., anticipated versus actual). If there is a
match, the path followed to arrive at the output point is followed through backpropagation, and the intervening nodes and paths are reinforced for that particular
output point.
Learning within a neural network requires feedback that allows the network
to compare the expected output value associated with a set of input data against
the conclusion reached by the neural network. This forms the back-propagation
illustrated in Figure 12.13. As the vectors of input values are applied to the system, the data propagates through the intermediate layers to the output. The
expected output is mapped onto the output vectors. Those elements in an output
391
Chapter 12
Figure 12.13: Neural network.
Shown are forwardpropagation of input and
back-propagation learning.
Input
Output
Reinforcement
Propagation
vector whose value matches the applied expected value are reinforced by backpropagation. Thus, the intermediate layers that contributed to the propagation of
data resulting in the correct (i.e., expected) output values are reinforced.
Reinforcement may be performed through increasing weights, of the nodes
involved in the propagation from the input data to the correct output or
conclusion.
This reinforcement of neural links increases the probability that the same
input values will result in the same propagation to the correct output values. Those
output values that did not match the expected output value are weakened, thereby
decreasing the probability that they would be applied again given the same set of
conditions.
12.3.8 Genetic Algorithms
GAs utilize a vector that represents a given condition or input stimuli and action.
Based on the success of the action taken, the input vector is modified. In the case
of GAs, however, the vector may undergo alteration, or genetic mutation, in a
392
Cognitive Research: Knowledge Representation and Learning
random fashion. Even though this may introduce potentially large vector sets,
GAs may introduce new solutions through the random mutation process. This
enables a system to generate new knowledge and evaluate its effectiveness by
using empirical data collected through the operational environment.
The initial implementations of GAs utilized a bit sequence as the representation
of the gene. The bit pattern essentially consists of two parts. One part forms a pattern that is used to match a given set of conditions. The second part forms a bit
pattern representing the action or actions to be taken as a result of the pattern
match.
As the process of matching the input part and performing the corresponding
actions is performed, two modifications of the “gene” occur. The first is a
strengthening of those genes that contributed to a successful set of actions (and a
weakening of those genes that did not). This serves to enhance the survival rate of
those genes that contributed positively, and effectively causes those genes that did
not contribute positively to die off (i.e., they are removed from the set of bit
sequences).
The second action is a random mutation of the bit sequence. This provides a
mechanism for generating new alternatives within the set of bit sequences. Thus,
not only will genes live or die off through a form of natural selection, but wholly
new strains can be introduced that may result in more efficient reasoning.
Implementation of GAs has continued to evolve to include the use of a vector
of multivalued data items. In effect, each item in the sequence is represented as a
discrete variable that can be modified as part of the process. See Chapter 7 for a
detailed discussion of GAs.
12.3.9 Simulation and Gaming
The application of simulation and game theory to machine learning provides a
method for developing and evaluating alternative actions in a given scenario,
thereby proposing new paths or actions. At a fundamental level, game-theoretic
learning utilizes a reinforcement mechanism similar to those previously discussed.
Unlike reinforcement and temporal learning, which require a comprehensive
model of states, however, game theory—due to its incorporation of strategy and
probability—supports learning in systems that are not well defined, or are changing. Game-theoretic learning utilizes game theory as a means for proposing
actions to be taken, given a particular situation.
Game-theoretic learning introduces some features that are unique to this
approach. The player chooses from a set of actions for any given state. The action
may be selected based on a numerical probability, random generation, or any
393
Chapter 12
other selection method. As the player applies the selected strategy, the success of
the strategy is recorded for future reference. Thus, as in the reinforcement and
temporal difference algorithms, there is a bias value applied to a set of actions for
a given situation. Chapter 15 provides a comprehensive discussion of the application of game theory to cognitive radio systems.
12.4 Implementation Considerations
The previous sections discussed different approaches to knowledge representation
as well as reasoning and learning approaches. This section discusses implementation and deployment considerations related to the use of this technology within a
radio system. Several areas are related to the technological requirements necessary to realize a cognitive radio that learns, and some of the key issues that need
to be addressed are related to operational and sociological issues that are raised
when a learning system is deployed.
12.4.1 Computational Requirements
As with any type of work, some effort must be expended in order to produce
the desired artifact or product. Cognitive radio systems must expend effort in
the form of computational resources and the power to drive them in order to
reach conclusions or decisions concerning the operation of the system. Similarly,
computational effort must be expended by a learning system in order to
monitor and analyze the results of the actions performed against the expected
outcome, perform analysis on the differences between the two, and generate
new or changed knowledge to be applied to the next iteration of the decision
process.
It can be argued that whereas the cognitive decision process is an integral part
of the operational behavior of the radio system, the learning process is not.
Consequently, relegating the learning process to non-critical performance times
would be a reasonable operational assumption. However, the primary driver is the
availability of computational resources that can be applied to the learning process
without affecting the mission-critical performance of the radio system.
12.4.2 Brittleness and Edge Conditions
Learning has the potential to overcome the brittleness of cognitive implementations when faced with a new situation or scenario. However, the resources and
time required to generate additional or new knowledge that can be applied to the
394
Cognitive Research: Knowledge Representation and Learning
situation may not be sufficient to result in a feasible alternative in a time frame
that can be applied to the situation. Nonetheless, the ability of the system to
extend its behavior provides unique opportunities for radio systems. Systems that
are to be deployed for long-duration missions in locations that cannot be easily
serviced would benefit from the capabilities afforded by learning algorithms.
12.4.3 Predictable Behavior
One of the critical aspects of any automated system is the tacit assumption that the
system is deterministic in its behavior. That is, given a set of inputs, the output, or
action taken by the system, can be determined. The problem that arises in any
complex system is that each step taken to reach a decision may be individually
predictable, but the end decision or action reached may not be what was expected
on the part of the human operating or interacting with the system. This trait of
individually predictable steps leading to an unexpected or unforeseen outcome is
one of the founding premises of chaos theory. The unpredictability of the endpoint
action is further exacerbated by the introduction of learning to the system.
Enabling learning within a system is a powerful two-edged sword. On the one
hand, it enables the radio system to adapt autonomously to new situations, change
operational characteristics, make operational decisions based on newly learned
knowledge, and provide more reliable and dependable service. On the other hand,
as systems learn, there is no guarantee that the observed data used by the learning
system to adapt or form new knowledge is accurate, nor is it assured that the
methodology applied to form some new chunk of knowledge is without flaw.
Just as in the case of humans, the ability to extend the knowledge and behavior of the system is not assured to be predictable. Yet, it is predictability that
underlies current methodologies and approaches to radio system certification.
Regulatory bodies, such as the US Federal Communications Commission (FCC),
are just beginning to consider the ramifications of a radio system that is largely or
exclusively software based and may change its operational characteristics through
changes in the software. Regulatory methods and approaches, as discussed in previous chapters, will be stretched, as cognitive radio technology becomes an integral part of radio systems, allowing them to autonomously change operational
characteristics. Finally, as learning technologies are integrated within cognitive
radio systems, the concept of radio system certification will again need to evolve
as cognitive radio systems gain the ability not only to change operational behavior
based on the operational environment but also to modify the method by which the
decisions to change operational behavior is reached.
395
Chapter 12
12.5 Summary
This chapter has reviewed several methods of knowledge representation and reasoning, approaches to computational learning, and possible applications of these
methods within the domain of cognitive radio systems. As shown, there is no single approach to knowledge representation and reasoning or learning that will
address all aspects of cognitive radio systems. The level of intelligence exhibited
by a cognitive radio system will be dynamic and varied, depending on the implementation platform, the computational resources available, and the specific mission of the system.
Also, the type of knowledge representation and reasoning system employed as
well as the type of knowledge may “prefer” a particular learning method. Radio
systems may therefore employ a particular representation mechanism, and hence a
specific learning implementation based on the operational constraints and objectives imposed. Some systems may host two or more methods, or may interact with
other radios in a networked fashion to implement a larger intelligent entity.
Learning within the context of a cognitive radio must encompass multiple
representation paradigms, integrate multiple reasoning mechanisms, provide
hybrid-learning mechanisms, and span multiple functional layers within the cognitive radio. This concept is illustrated in Figure 12.14.
The cognitive radio can be viewed through three essential aspects or
perspectives:
●
●
●
At the lowest layer is the physical set of hardware. This provides the physical
infrastructure for the radio system.
The infrastructure aspect provides an abstraction layer that implements a logical set of interfaces to manage, control, configure, and operate the physical
components and provide software views of these resources to the waveform
implementation. This layer also manages the suite of components that implement a waveform.
At the application and services layer, waveforms and system services are
viewed and manipulated as a logical entity.
A fourth perspective is the user’s aspect. From that viewpoint, the user is either
interacting with the physical radio (e.g., power up, diagnostic checks, etc.) or
interacting with an instantiation of a waveform as a logical entity.
As the right-hand side of Figure 12.14 illustrates, knowledge representation,
reasoning, and learning must be implemented within and across each of these
396
User
Application and
Services
Waveform
«real ize»
Plan Generation
and Execution
Service
1..*
Software
Infrastructure
Port
connects Via
Resource
0..*
Mission Objectives
and Goals
Spatial/Temporal
Reasoning
«real ize»
controls
has Dependency
User
Model
SWComponent
1
issues
0..*
Radio User
1
Property
LogicalDevice
«real ize»
provides
Interface
requires
Event
1
is Allocated To
Hardware
Capacity
1..*
0..*
Resource
Constraint Engine
Configuration
Action Logic
Internal/External
Safety Constraints
1
has Dependency
Radio
PhysicalElement
Processor
I/O Device
NIC
AD-DA
GPP
DSP
RF
FPGA
Antenna
Physical/Sensory
Layer
397
AppComponent
operates
Reactive/Reflexive
Layer
1..*
Representation, Reasoning, and Learning
has Dependency
SWApplication
Reasoning
Layer
«real ize»
Device
Characteristics
Environmental
Sensors
Operational
Constraints
Figure 12.14: Multiple learning methods. Learning and reasoning span multiple architectural layers within the cognitive radio.
Chapter 12
layers, yielding a hybrid system. Physical hardware will become smarter with
low-level implementation of simple learning algorithms to adaptively modify the
power usage or RF components. The software infrastructure will provide a basic
reflexive behavioral and learning capability that provides, among other capabilities, safety limitations and operational constraints. At the upper levels of the cognitive radio, reasoning and learning become more abstract, addressing mission
requirements and user needs.
Future research and development of computational learning applications for
radio systems will need to address the hybrid nature of reasoning and learning
within cognitive radio systems.
References
[1] J. Mitola III and G.Q. Maguire Jr., “Cognitive Radio: Making Software Radios
More Personal,” IEEE Personal Communications, Vol. 6, No. 4, August 1999,
pp. 13–18.
[2] J. Wang, D. Brady, K. Baclawski and M. Kokar, “The Use of Ontologies for
the Self-Awareness of the Communications Nodes,” in Proceedings of the
Software Defined Radio Technical Conference—SDR03, Orlando, FL,
November 2003.
[3] J. Wang, M. Kokar, K. Baclawski and D. Brady, “Achieving Self-Awareness of SD
Nodes Through Ontology-Based Reasoning and Reflection,” in Proceedings of the
Software Defined Radio Technical Conference—SDR04, Phoenix, AZ, November
2004.
[4] M. Kokar, D. Brady and K. Baclawski, “Roles of Ontologies in Cognitive Radios,”
in Cognitive Radio Technology, B. Fette (Ed.), Elsevier, 2006, pp. 401–434.
[5] M. Minsky, “A Framework for Representing Knowledge,” Massachusetts Institute
of Technology, Technical Report, UMI Order Number: AIM-306, 1974.
[6] S. Bechhofer, et al., “OWL Web Ontology Language Reference,”
http://www.w3.org/TR/owl-ref/, M. Dean, G. Schreiber (Eds.), February 2004.
[7] K. Baclawski, D. Brady and M. Kokar, “Achieving Dynamic Interoperability of
Communication at the Data Link Layer Through Ontology Based Reasoning,”
in Proceedings of the Software Defined Radio Technical Conference—SDR05,
Garden Grove, CA, November 2005.
[8] R. Schank, Dynamic Memory: A Theory of Learning in Computers and People,
Cambridge University Press, Cambridge, England, 1982.
[9] J. Kolodner, “Improving Human Decision Making Through Case-Based Decision
Aiding,” The AI Magazine, Vol. 12, No. 2, 1991, pp. 52–68.
398
Cognitive Research: Knowledge Representation and Learning
[10] C. Forgy and J. McDermott, “OPS, A Domain-Independent Production System
Language,” International Joint Conference on Artificial Intelligence, 1977,
pp. 933–939.
[11] J.F. Allen, “Maintaining Knowledge about Temporal Intervals,” Communications of
the ACM, Vol. 26, No. 11, November 1983, pp. 832–843.
[12] V. Kovarik and A. Gonzalez, “An Interval-Based Temporal Algebra Based on
Binary Encoding of Point Relations,” International Journal of Intelligent Systems,
Vol. 15, No. 6, June 2000, pp. 495–523.
[13] J.F. Allen, “Time and Time Again: The Many Ways to Represent time,” International Journal of Intelligent Systems, Vol. 6, No. 4, July 1991, pp. 341–356.
[14] http://en.wikipedia.ord/wiki/Ivan_Pavlov
[15] A. Tversky, “Features of Similarity,” Psychological Review, Vol. 84, No. 4, July
1977, pp. 327–352.
[16] R. Sutton and A.G. Barto, Reinforcement Learning: An Introduction, MIT Press,
Cambridge, MA, 1998.
[17] R.S. Sutton, “Learning to Predict by the Methods of Temporal Differences,”
Machine Learning, Vol. 3, No. 1, 1988, pp. 9–44.
399
This page intentionally left blank
CHAPTER 13
Roles of Ontologies in Cognitive Radios
Mieczyslaw M. Kokar, David Brady and Kenneth Baclawski
Northeastern University
Boston, MA, USA
13.1 Introduction to Ontology-Based Radio
This chapter discusses the role of knowledge representation (ontologies) in cognitive radio. The emphasis is on those capabilities of cognitive radio that are practical and amenable to ontological treatment. After a brief introduction to ontologies,
this chapter develops an ontology for cognitive radio functionality that is organized by using the same layers as communication networks. Specific examples of
the use of ontologies in cognitive radio are then developed in some detail. The
examples show how ontologies can be the basis for achieving interoperability at
the physical (PHY) and data link (DL) layers. This chapter then outlines some of
the major research issues for ontology-based cognitive radio.
13.2 Knowledge-Intensive Characteristics of Cognitive Radio
The term cognitive radio is interpreted differently by different people. This chapter uses an interpretation that includes a cognitive agent, as well as standard radio
functionality. Toward this aim we first discuss the features that a cognitive agent is
expected to have, and then propose some features that a cognitive radio could have.
To put the discussion in context, a cognitive radio is viewed as being part of a
larger functionality, that is, of an intelligent agent or intelligent personal digital
assistant (PDA) that can support a mobile user. Such a PDA would not only have
to advise the user, but it would also have to be connected essentially all the time.
Although there are many definitions of a cognitive agent, all of them revolve
401
Chapter 13
around similar ideas. For instance, the Defense Advanced Research Projects
Agency (DARPA) [1] provides the following definition of a cognitive system:
A cognitive system is one that:
●
●
●
●
●
can reason, using substantial amounts of appropriately represented knowledge
can learn from its experience so that it performs better tomorrow than it did
today
can explain itself and be told what to do
can be aware of its own capabilities and reflect on its own behavior
can respond robustly to surprise.
The functionalities of a cognitive agent are often viewed according to Boyd’s
so-called OODA (Observe, Orient, Decide, Act) loop [2]. This approach is
especially popular in the information fusion community where the OODA loop is
used to model human behavior, which then serves as a pattern to be followed by
a fusion system. In the intelligent control community, this loop is presented in a
somewhat simpler form, called the perception-reasoning-action triad [3], used to
represent an intelligent controller.
Following roughly the same line of reasoning, this chapter identifies the following basic functionalities that a cognitive radio should include:
●
●
●
●
●
●
●
●
information collection and fusion;
self-awareness;
awareness of constraints and requirements;
query by user, self or other radio;
command execution;
dynamic interoperability at any stack layer;
situation awareness and advise;
negotiation for resources.
Each of these capabilities is discussed in this chapter from two points of view:
(1) the user’s point of view and (2) the radio’s point of view. This reasoning is
illustrated and explained in Figure 13.1.
As shown in Figure 13.1, multiple conversations are taking place at several
layers. At a higher layer, the two speakers pose and respond to queries through
their radios; simultaneously, at a lower layer the radios converse in a similar
402
Roles of Ontologies in Cognitive Radios
User Application Layer “Cover Area B3”
Sensor Layer
Situation Awareness
Device Layer “What is your multipath Structure?”
…, delayi, ampi, phasei,…,
Figure 13.1: Three layers of cognitive radio conversation: user layer, sensor layer, and device
layer. Cognitive radios permit conversations at all three layers. Conversation at user layer is
a standard feature of all radios. Communication at the sensor layer is required for selfawareness, situation awareness, and information collection. Communication at the radio
layer enables resource negotiation, query, and response.
fashion. In this case, one radio is trying to establish the multipath characteristics
observed by the other with the intent of improving performance. In order to
achieve this goal, the radios need to monitor the communication (indicated in the
figure as dashed arrows between the radios and their users) as well as possibly
some environmental characteristics, for instance, obtained from other sensors
(indicated by the arrows from the radios to the “ancient astronomer” icon). The
notion of simultaneous communications at different levels may be applied to each
layer in the protocol stack, as discussed in Section 13.2.8. These conversations
will be interrelated by the requirement that the “layer below” is supporting the
“layer above”.
13.2.1 Knowledge of Constraints and Requirements
At the base of any cognitive activity is knowledge. A cognitive agent must possess
knowledge and must be able to make use of the knowledge for deriving its decisions and actions. Any software system contains some knowledge, but the real
issue is how knowledge can be represented so it can be used in the most flexible
way. This issue is related to the requirement that a cognitive system “respond
403
Chapter 13
robustly to surprise.” When knowledge is encoded in a purely procedural way, the
use of that knowledge must be prescribed (encoded) during system development.
Thus, it is the responsibility of the system developer to not only encode all the
necessary procedures, but also to encode the invocations of the procedures. In
contrast to this approach, the declarative programming approach requires only
that knowledge fragments be represented in the system’s data structures (e.g.,
rules); the selection of the knowledge is left to the system. Such decisions are
made by the system at run-time based upon pattern matching; that is, specific
knowledge fragments are used when specific patterns are recognized by the system. This approach gives more flexibility to the use of stored knowledge and at
least partially satisfies the requirement of the system responding to surprise.
13.2.2 Information Collection and Fusion
In addition to knowledge encoded in the cognitive agent, the agent must also have
access to current information. Whereas knowledge is encoded by either the system designer or the user, the current information is collected from various sources
at run-time. For instance, a PDA could have access to a global positioning system
(GPS) receiver and thus obtain information of its current location. Additionally, a
PDA could (potentially) have access to traffic reports, including the status of roads
and bridges, as well as accident reports, weather conditions such as blizzards and
black ice, and other pertinent details. Based on such information, the PDA could
advise the user on efficient routing. Some of this information could be collected
by subscribing to a service, whereas other information would require explicit
requests. Because such advice would depend on various types of information from
multiple sources, the information would have to be integrated and fused appropriately in order to be useful to the user.
13.2.3 Situation Awareness and Advice
Having knowledge and current information stored in the PDA’s memory does not
automatically guarantee that the PDA is aware of the current situation of the user
(and, in effect, the PDA). According to Endsley and Garland [4], “Situation
Awareness is the perception of the elements in the environment within a volume
of time and space, the comprehension of their meaning, and the projection of their
status in the near future.” In other words, in order to understand a situation, the
agent needs to know not only about all the objects of interest, but also about their
relations to other objects and also possible future states of the objects and the relations. This, in turn, requires that knowledge of models of objects, including their
404
Roles of Ontologies in Cognitive Radios
dynamics, be known, so that future states could be predicted and rules for determining the relationships could be derived. With respect to the PDA’s task as a
travel advisor, an example of a relevant relation might be “on-the-path-from-to,”
meaning that a specific geographic feature, such as a bridge or a river, is on the
path that leads “from-to” specific destinations.
13.2.4 Self-awareness
To plan and schedule a task for execution, a PDA must know whether the task is
within its own capabilities. It needs to understand what it does and does not know,
as well as the limits of its capabilities. This is referred to as self-awareness. For
instance, the radio should know its current performance, such as bit error rate
(BER), signal-to-interference and noise ratio (SINR), multipath, and others. In a
more advanced case, the agent might need to reflect on its previous actions and
their results. For instance, for the radio to assess its travel speed a fortnight ago
between locations A and B, it might be able to extract parameters from its log file
and do the calculation. For the radio to decide whether it should search for the
specific entries in the log and then perform appropriate calculations (or simply
guess), it needs to know the effort required to perform such a task and the required
accuracy of the estimate to its current task.
13.2.5 Query by User, Self, or Other Radio
Similar to humans, intelligent agents need to be able to answer queries from their
users and other agents. Moreover, they should be able to query the user and other
agents about relevant information whenever they cannot infer it from their own
knowledge. For agents to formulate such queries, they must understand a common
language, formally defined in syntax and semantics. Moreover, to formulate
queries, agents must have a planning capability. Finally, they should be able to
wait for an answer and incorporate the answer into their own knowledge structures. For instance, the PDA might ask the user, “Where do you go next?” or ask
another agent, “What is the temperature at your location?” Another radio could
ask, “What is your multipath structure?” or “What is your BER?”
13.2.6 Query Responsiveness and Command Execution
If two humans agree on a plan to perform a specific task, they then (normally)
execute the plan by performing the actions they agreed to. For instance, if two
persons are driving around the neighborhood searching for a runaway dog, they
405
Chapter 13
communicate over cell phones and decide who will search which street and when.
The implementation of such a feature in a software radio is not a simple matter.
For instance, if two radios decide that one of them should adjust its communication protocol after exchanging information about the multipath structure, by selecting a different modulation scheme, the radio must be able to understand how to
implement such a command. A simple way would seem to be to provide a number
of procedures and invoke one of them using an if–then–else programming construct. However, this would limit the flexibility of reacting to different multipath
structures to only the procedures encoded at the development time of the software
radio. Moreover, the invocation of the procedures would have to be decided at
development time. At the other extreme, one might envision an approach in which
code is automatically generated at run-time and then executed. A middle ground
solution would be to use the declarative programming approach in which knowledge fragments are developed and represented as rules at design time and then the
run-time system selects, combines, and executes the rules as needed. This
approach would provide more flexibility in terms of possible behaviors in
response to various external conditions.
13.2.7 Negotiation for Resources
The human activity of planning involves, among others, negotiation for resources.
For instance, if two people decide that one of them will drive to a grocery store to
get drinks and the other will drive to the hardware store to buy a ladder, and if
they have a sedan and a SUV at their disposal, it is likely that through a rational
negotiation process they will use the SUV for the hardware store and the sedan
for the grocery store. Similarly, resource negotiation is very useful in software
radios. Negotiation of the most important resource, the spectrum, is being investigated in the DARPA NeXt Generation (XG) program [5]. According to the XG
vision, software radios would be able to request unused spectrum that is allocated
to another radio. This approach should lead to the achievement of a higher utilization of the spectrum resource.
13.2.8 Dynamic Interoperability at Any Stack Layer
The flexibility of ontology-based radios (OBRs) postulated in the above discussion can be extended to protocols at any layer of the protocol stack. This extension would enable intercommunication and negotiation among layers. Query and
response among layers invites the possibility of cross-layer optimization of
406
Roles of Ontologies in Cognitive Radios
communication efficiency while maintaining the strict functional division
between layers. For example, the medium access control (MAC) layer, which
handles channel access and routing, might query the local PHY layer about the
residual error in an equalizer, with the goal of preempting an outage. If the
reported error is sufficiently large, the MAC might avoid an outage by seeking
an alternative channel and route to the destination. As another example, MAC
and transport layers at different nodes may also use query and response to reduce
the frequency of broadcasting routing tables. Through the process of query and
response, MAC layers could also provide a much more meaningful routing
metric than hop count or aggregate delay because much more PHY layer information would be available. DL layers, which manage automatic repeat request
(ARQ) protocols, may avoid some of the rigidness of fixed protocols such as
selective repeat go-back-n or stop-and-wait. Through the use of reasoning,
ARQs may be tailored on the fly for individual links as the nodes learn about the
link performance.
This chapter focuses on ontological presentations for the PHY and DL layers
because these layers are the ones that were implemented in hardware or firmware
prior to the introduction of software-defined radio. Higher layers are more likely
to be already implemented in software. Furthermore, the lower layers present
challenges that are unique to radio communication.
13.3 Ontologies and Their Roles in Cognitive Radio
13.3.1 Introduction
An ontology is an explicit mechanism for capturing the basic terminology and
knowledge (the concepts) of a domain of interest as well as the relationships
among the concepts [6]. Ontologies are an increasingly important mechanism for
the integration of disparate software systems. Indeed, a shared ontology is a fundamental prerequisite for meaningful communication between systems. The
ontology can be hard-coded via shared data formats, database schemas, and procedures, but there are significant advantages for expressing the shared ontology by
using a formal declarative ontology language that is either difficult to achieve or
even impossible without it. The advantages include support for interoperability,
flexible querying, run-time modifiability, validation against specifications, and
consistency checking.
In the case of software-defined radio, ontologies offer the additional advantage of self-awareness: communication nodes can understand their own structure
and can modify their functioning at run-time. Furthermore, nodes can query the
407
Chapter 13
capabilities and current state of other nodes, allowing them to modify the processing of packets during a communication session both at the source and the destination. The ontology specifies not only the structure of communication packets but
also the processing of those packets according to the communication protocol.
The use of ontologies adds flexibility, inferencing, and reasoning features that are
not available with ad hoc data structures or database schemas [7, 8].
Systems that do not initially share an ontology might still be able to interoperate by mapping or merging ontologies in order to synthesize a shared ontology. It
may be possible to construct systems that not only use ontologies but also modify
them or even learn them dynamically.
Basics
An ontology specifies the concepts of a domain, attributes of the concepts, and
relationships among the concepts. Each concept is expressed by using a class,
which may be interpreted as a set of things. Anything that belongs to a class is
called an instance of that class. Waveforms, packets, and symbol alphabets are all
fundamental concepts in radio communication. In the ontology for software radio,
these concepts are expressed as classes. For example, Waveform is a class whose
instances are particular waveforms. Classes are organized into a hierarchy of
classes by the subclass relationship. For example, binary phase shift keying
(BPSK) and quadrature phase shift keying (QPSK) are both special cases of
M-ary phase shift keying (MPSK). This is expressed by specifying that BPSK
and QPSK are subclasses of MPSK.
An attribute is a property that something has, such as the number of symbols
in an alphabet or the carrier frequency of a waveform. An attribute is a characteristic of a single entity, where that characteristic is a data value such as a number.
A relationship is an association among various entities. For example, a waveform
is used to represent a sequence of symbols from an alphabet. This is expressed
by linking the waveform to the symbol sequence. An ontology will generally
have many different kinds of attributes and relationships. The number of symbols
in an alphabet might be called numberOfSymbols, and the relationship of a
waveform with the sequence of symbols it represents might be called the
usedToRepresent relationship. The term property is used for either an attribute
or a relationship. As with classes, one kind of property may be regarded as a set of
instances, called facts. For example, when a particular waveform w is being used
to represent a particular sequence S, this fact is the triple (w, usedToRepresent, S).
Properties can be organized in a hierarchy by the subproperty relationship. For
408
Roles of Ontologies in Cognitive Radios
example, the usedToRepresent property is a subproperty of the more generic
used property.
Ontology Languages
The rapid expansion of the World Wide Web (WWW) has had a profound impact
on communication. One consequence of this expansion is the emergence of the
eXtensible Markup Language (XML) as the most commonly supported format
for data interchange [9]. This trend has also affected ontology and knowledge
representation languages. In order to be interoperable, it is becoming essential that
such languages be web based and expressible in XML. Being web based means
that the ontology is concerned with web resources. A web resource is identified
by a universal resource indicator (URI). In other words, anything being described
by a web-based ontology is a web resource.
There are three major web-based ontology languages: XML Topic Maps (XTM)
[10], the Resource Description Framework (RDF) [11] and the Web Ontology
Language (OWL) [12]. XTM is an International Organization for Standardization
(ISO) standard (ISO13250), whereas RDF and OWL are standards of the World
Wide Web Consortium (W3C). XTM allows one to specify relationships among
any number of web resources. By contrast, RDF and OWL restrict all relationships to be binary. Restricting relationships to binary simplifies the language
and processors, but it makes it much more awkward to deal with relationships
that involve more than two resources. It also restricts the portability of knowledge
structures to hardware platforms other than those on which the structure was
created.
When one restricts relationships to be binary, all facts are triples, consisting
of the two entities being related (called the subject and object) and the relationship between them (called the predicate). RDF is an XML-based language for
representing triples. The RDF Schema (RDFS) language is an extension of RDF
that allows one to specify subclass and subproperty relationships. It also allows
one to specify the domain and range of a property.
The OWL language has three levels: OWL Lite, OWL-DL, and OWL Full.
They differ in the constructs that are allowed, with Lite being the most restrictive
and Full being the least restrictive. OWL adds many new capabilities to RDF and
RDFS, such as cardinality constraints, disjointness constraints, enumerations, and
inverse properties. However, the most significant new feature is the ability to construct classes from other classes, such as defining a class to be the intersection or
union of two or more other classes. For example, a BPSK is the special case of
409
Chapter 13
MPSK for which there are exactly two symbols in the waveform alphabet (each
one being 180 degrees away from the other). Class constructors are the basis for a
form of reasoning known as description logic [13].
OWL Lite differs from OWL-DL in the class constructors that are allowed.
For example, in OWL-DL, one can specify the complement of a class (i.e., all
instances that are not in the class), but this specification is not allowed in OWL
Lite. Although OWL-DL allows all class constructors, it does not allow one to
cross metalevels. For example, in OWL-DL a class cannot be an instance of
another class. By contrast, in OWL Full, a class can be an instance of another
class, which itself is an instance of yet another class, and so on to any number of
levels. Although OWL Full is a very rich ontology language, it is still not as rich
as arbitrary first order predicate logic. Table 13.1 summarizes the features and
differences between the various web-based ontology languages.
Table 13.1: Existing web-based ontology languages and their characteristics. The last column describes the level of reasoning as well as the level of processing complexity as a
function of the number of triples.
Language/
organization
Features
Reasoning/complexity
XTM/ISO
Higher order relationships
None/linear
RDF/W3C
Binary relationships
None/linear
RDFS/W3C
RDF plus subclass, subproperty,
domain, and range
Subsumption/polynomial
OWL Lite/
W3C
RDFS plus some class
constructors, but no crossing
of metalevels
Limited form of description
logic/exponential
OWL-DL/
W3C
All class constructors, but no
crossing of metalevels
General description logic/
decidable
OWL Full/
W3C
No restrictions
Limited form of first-order
predicate logic/undecidable
One might think that the richer the ontology language the better. However,
richer ontology languages are also more difficult to process. RDF and RDFS are
relatively easy to process. The time to process RDF is linear in the number of
triples. RDFS is a little more difficult, being polynomial in the number of triples
and NP-complete in the size of the query in the worst case. OWL Lite is much
more difficult, requiring exponential time in the number of triples for the worst
410
Roles of Ontologies in Cognitive Radios
case. OWL-DL is decidable, meaning that the processing will finish in a finite
amount of time, but the amount of time can be more than exponential. Finally,
OWL Full is undecidable, which means that ontologies cannot be processed, and
some queries cannot be answered, in a finite amount of time.
Querying
Given a database, one can extract information by using a query. Many query languages have been proposed for XTM, RDF, and OWL. The query language that
has been proposed for RDF is called RDQL [14]. As is the case with most query
languages, RDQL is syntactically and semantically very similar to the Structured
Query Language (SQL) for relational database systems. RDQL differs primarily
by allowing one to specify patterns. A pattern is a fact in which some of the
components can be variables.
For example, the query in Figure 13.2 retrieves all fields of all DL frames. The
query patterns in Figure 13.2 specify both variables, such as ?x, and constants,
such as http://ontoradio.org/contains. The constants are the URIs of web
resources. The ontoradio.org domain is the OBR domain where the OBR ontology
resources are defined. The query language that has been proposed for OWL is
called OWL-QL [15].
1.
2.
3.
4.
SELECT ?x
WHERE (?x, <http://www.w3.org/1999/02/22-rdf-syntax-ns#type>,
<http://ontoradio.org/2005/datalink#DataLinkFrame>)
AND (?y, <http://www.w3.org/1999/02/22-rdf-syntax-ns#type>,
<http://ontoradio.org/2005/datalink#DataLinkField>)
AND (?x, <http://ontoradio.org/2005/datalink#contains>,?y)
Figure 13.2: An example of RDQL, a query language for RDF. The first line enables the
retrieval of attributes of variable ?x that match patterns. Patterns are expressed as triples
(subject, predicate, object) following the WHERE delimiter. The second line limits all ?x to be
of type DataLinkFrame. The third line describes those attributes that are of interest (those ?y
of type DataLinkField). The last line restricts the objects to be retrieved to be DataLinkFields
contained in DataLinkFrames.
The RDQL and OWL-QL query languages differ from database query languages in several important ways. One important difference is that they are web
based. Whereas databases are generally restricted to a single server or at least one
site, RDF and OWL effectively regard the entire web as being a single database.
411
Chapter 13
OWL-QL has the additional feature of specifying a protocol for query requests
and answers to support its use by web services.
The most important difference between database queries and query languages
such as RDQL and OWL-QL is their support for reasoning. In addition to facts
that have been explicitly asserted, a query can also retrieve facts that have been
inferred. It is for this reason that RDF and OWL databases are said to be knowledge bases. A knowledge base is the set of all currently known or inferred facts in
a particular context. The reasoning capability of OWL is especially powerful. We
elaborate on this feature in section Reasoning.
Radio communication introduces an additional requirement on query languages. Unlike databases, which have explicitly stored data, and knowledge bases,
which have a combination of explicitly asserted facts and inferred facts, an OBR
knowledge base includes data that is embedded in the software that implements
the communication protocols. Extracting such data requires a new software capability known as self-awareness, or reflection.
Reflection is a property that enables software to understand its own run-time
structure. Reflection is a key feature of any software system that is expected to
respond to unanticipated queries. Run-time structure is significantly different from
program structure, which is an artifact of the language itself, because it also is a
function of the compiler and the run-time platform. Run-time structure includes
memory pointers for variables, for their subfields, for methods, and for procedures. This is essential for the software to be able to answer queries and execute
methods dynamically.
Reasoning
One of the important features of ontology languages that distinguishes them from
databases is the ability to make logic deductions. In other words, one can reason
about the information in a knowledge base. A fact is deduced if one can infer that
it is true even though the fact has never been explicitly asserted to be true. One of
the most important examples of deduction is subsumption. For example, if one
knows that an analog signal uses BPSK, then one can deduce that it is also of type
MPSK. In general, whenever something is an instance of a subclass, then it is also
an instance of all superclasses of the subclass. Subsumption is the basis for reasoning in description logic. For example, all features and axioms applicable to
MPSK signaling also apply to BPSK signaling.
While subsumption reasoning is useful, it is not sufficient for all reasoning tasks.
When reasoning involves several linked facts, one cannot express the inference
412
Roles of Ontologies in Cognitive Radios
using subsumption alone. When database records are linked by common attributes, they are said to be joined. To express reasoning involving joins of facts, it is
necessary to introduce rules. A rule is knowledge in the form of an if-then statement. If a hypothesis holds, then a conclusion must also hold. The hypothesis is
also known as the antecedent, and the conclusion is also known as the consequent.
The rule language that has been proposed for OWL is the Semantic Web Rule
Language (SWRL) [16]. An example of a SWRL rule is shown in Section 13.4.2.
13.3.2 Role of Ontology in Knowledge-Intensive Applications
Two-way radio communication introduces a number of challenges not shared by
most other ontology-based applications:
1. Real-time processing demands higher performance for inference and reasoning
than an interactive application.
2. The knowledge base of a node includes state information that is continually
varying, in contrast with the static knowledge bases required by most reasoning systems.
3. The facts are not explicitly stored in a knowledge base but rather are embedded
in the software that implements the communication protocol [2].
4. The radios may not have access to the WWW or the Semantic Web for broad
support.
5. Most radios are continually moving, and the link performance is bandwidth
restrictive and time varying.
If these challenges can be overcome, ontologies can play a number of important
roles in radio communication, such as the following:
1. Interoperability: Radios can use ontologies to deduce important information,
such as the protocol being used by other radios.
2. Flexible querying: Information, such as multipath structures, can be queried.
Furthermore, such queries can be answered without having any explicit preprogrammed monitoring capability.
3. Run-time modifiability: Protocols, packet structures, and even waveforms can
be modified at run-time in response to environmental conditions and application requirements.
413
Chapter 13
4. Validation: Formalization allows one to check the consistency of protocols and
to validate the correctness of algorithms that implement the protocols.
5. Self-awareness: Communication nodes can understand their own structure and
modify their functioning at run-time based on this understanding.
13.4 A Layered Ontology and Reference Model
This section discusses two (partial) ontologies, for the PHY layer and for the DL
layer. These ontologies are then used for the discussion of the realization of the
cognitive aspects discussed in Section 13.5.
13.4.1 Physical Layer Ontology
A piece of the ontology for the PHY layer is shown in Figure 13.3. It is represented
in the Unified Modeling Language (UML) notation. UML is used primarily in software engineering to represent software and systems. The boxes in UML represent
classes, and the connecting lines represent associations. UML classes correspond
to the classes used in object-oriented programming languages such as C, Java,
or C#. For programming languages such as C, which are not object oriented,
classes are implemented by using structs. Associations can be implemented in
a variety of ways, depending on the kind of association. In OWL, associations
are called properties.
Some associations have a predefined meaning. For instance, the hollow arrow
at the end of the line in Figure 13.3 represents the subclass relation. In the figure,
M-FSK, QAM, and M-PSK are all subclasses of Alphabet. The subclass relation
is supported by all object-oriented programming languages. In Java and C# one
specifies that a class is a subclass of another by using the “extends” keyword.
C is more succinct, specifying a subclass by using the colon character (“:”).
Diamonds represent aggregation. Thus Alphabet is an aggregate of symbols
(instances of class Symbol).
Aggregations are normally implemented by using arrays or linked lists, but
other implementations are also used, depending on the number of elements in the
aggregate and whether the aggregation has a variable number of elements.
Because a particular Alphabet has a fixed number of Symbol instances, it would
be implemented as an array. Non-object-oriented programming languages, such as
C, implement the subclass relation by using aggregation. Other associations are
identified either by the names placed in the middle of the association line, such
as encodes, or by the roles that elements of the related classes play (roles are
414
Roles of Ontologies in Cognitive Radios
M-FSK
QAM
M-PSK
Alphabet
Symbol Sequence
Length
Symbol
1...*
theSymbolSequence
Random Data
1
Encodes
Header
1
Packet
1
Training Data
theSignal
Noise
Equalizer
equalizerError
equalizerLength
Signal
theMultipath
Multipath
rmsDelay
excessDelay
Hopping Frequencies
FHSS
Sample
NoSS
Spreading Type
Spreading Sequence
DSSS
Figure 13.3: A partial ontology for the PHY layer in UML [8]. The boxes denote classes,
and the lines connecting them represent associations. The hollow arrowhead denotes
the subclass relation, and the diamond denotes aggregation.
attached to the line ends and are distinguished by the “” symbol). So the
association encodes in Figure 13.3 has two roles, theSymbolSequence
and theSignal. Associations such as these are implemented by using pointers or
references to the associated object. In some cases, associations have a prespecified
multiplicity. For instance, an Alphabet must have at least one instance of Symbol.
Multiplicity specifications affect how an association can be implemented. Because
an Alphabet can have more than one instance of Symbol, it cannot be implemented using a single pointer or reference to a Symbol.
The ontology consists of two groups of classes. The upper part represents
classes and properties related to symbols. We can see that Packet consists of one
Header, one Training Data, and one Random Data. Each of them is an instance of
Symbol Sequence.
415
Chapter 13
The bottom part of Figure 13.3 represents classes and properties related to the
signal domain. The main association between the two parts is encodes. It relates a
Symbol Sequence to a Signal. Signal is an aggregate of instances of Sample. A
signal can be modulated in different ways. This ontology example shows two of
them, DSSS and FHSS. Another connection between the symbol and the signal
domain occurs through Equalizer, which estimates the Multipath using the
Training Sequence and then uses the result for signal interpretation.
13.4.2 Data Link Layer Ontology
The DL layer is responsible for transmitting frames and for error detection and
correction in communication links. There are many DL protocols. Each protocol
specifies the frame types and structure as well as how the communication link
is controlled. Many of the DL protocols specialize and extend other protocols.
Consequently, the DL protocols form a hierarchy as shown in Figure 13.4.
Because of the diversity and complexity of DL protocols, not all of the details of
the DL ontology are shown in the figure.
Data Link Protocol
Star LAN
CAN
SDLC
ATM
HDLC
SS7
LAPB
Local Talk
LAPD
WiFi
ADCCP
Figure 13.4: A partial hierarchy of data link protocols in UML.
There are a large variety of frame types. The names and semantics of the
frame types depend on the protocol, but there is considerable overlap among the
protocols. Consequently, although each protocol has its own frame type hierarchy,
the frame types in different hierarchies can be related by the OWL sameAs relationship. The frame type hierarchy for High-Level Data Link Control (HDLC)
protocols is shown in Figure 13.5 and the hierarchy for WiFi® protocols is shown
in Figure 13.6.
416
Roles of Ontologies in Cognitive Radios
HDLC Frame
Unnumbered
SNRM
SABM
DISC
Supervisory
Information
UA
RR
RNR
REJ
SREJ
Figure 13.5: HDLC frame hierarchy.
WiFi Frame
Data Frame
Authentication
Management
Association
Request
Deauthentication
Association
Response
Disassociation
Control
Reassociation
Response
Reassociation
Request
Probe
Request
Probe
Probe
Response
Beacon
Figure 13.6: WiFi frame hierarchy.
A frame consists of a sequence of fields. The frame structure is defined by the
order of the field types, the number of bits allowed in each field, and the values
(bit sequences) allowed in each field. The HDLC protocol has the field types
shown in Figure 13.7.
The WiFi protocol has many of the same field types, except for changes in terminology. For example, the WiFi Frame Control Field is called the Control Field
in the OWL syntax. Similarly, the WiFi Checksum Field is called the CRC
Field in the OWL syntax. In the OWL syntax, this is written as: Frame Control
Field owl:sameAs Control Field, and Checksum Field owl:sameAs CRC Field.
The DL classes are related to one another as shown in Figure 13.8.
A DL frame contains an ordered sequence of fields, each of which has a size
(in bits) and a value. A DL field belongs to a protocol and also has a mode. For
example, the HDLC protocol has three operational modes: NR (Normal
417
Chapter 13
HDLC Field
Address
Control Field
Opening
Information
Field
Sequence
Flag
Duration
CRC
Closing
Figure 13.7: HDLC field hierarchy.
Data Link Frame
Mode
Protocol
{Ordered}
Data Link Field
Data Link Mode
size: Nonnegative Integer
value: Bit Sequence
Data Link Protocol
Figure 13.8: DL layer ontology.
Response), AR (Asynchronous Response), and AB (Asynchronous Balanced).
These modes are not shown in Figure 13.8.
In addition to the hierarchies and relationships already described, the DL
ontology specifies a large number of rules that constrain DL fields within the same
frame and in related frames. Examples of such rules shown informally (and implemented in Figures 13.9 and 13.10) include:
1. If a frame belongs to the HDLC protocol, then its opening flag field has value
0 7E (i.e., the bit sequence 01111110).
2. If a frame belongs to the SDLC protocol, then the address field has 8 bits.
3. If a frame belongs to the SS7 protocol, then the address field has 0 bits.
4. If a frame belongs to the HDLC protocol, then the control field has 8 or 16 bits.
5. If a frame belongs to the HDLC protocol and the mode is available resource
map (ARM), then the address field has 0 bits.
6. If a frame belongs to the WiFi protocol then the first two bits of the control
field are zeroes. This subfield represents the version number.
418
Roles of Ontologies in Cognitive Radios
1 <owl:Class>
2
<owl:intersectionOf rdf:parseType="Collection">
3
<owl:Class rdf:about="#OpeningFlagField"/>
4
<owl:Restriction>
5
<owl:onProperty rdf:resource="#containedIn"/>
6
<owl:allValuesFrom>
7
<owl:Restriction>
8
<owl:onProperty rdf:resource="#protocol"/>
9
<owl:allValuesFrom rdf:resource="#HDLCProtocol"/>
10
</owl:Restriction>
11
</owl:allValuesFrom>
12
</owl:Restriction>
13
</owl:intersectionOf>
14
<rdfs:subClassOf>
15
<owl:Restriction>
16
<owl:onProperty rdf:resource="#value"/>
17
<owl:hasValue rdf:datatype="&xsd;hexBinary">7E</owl:hasValue>
18
</owl:Restriction>
19
</rdfs:subClassOf>
20 </owl:Class>
Figure 13.9: An OWL implementation of the DL layer protocol, which categorizes the
opening flag field of DL layer frames.
All of these rules can be represented either in SWRL or in OWL. For instance, the
first rule may be represented in OWL, as shown in Figure 13.9.
In XML, element tags, attribute names and attribute values can belong to various namespaces. The namespace of a tag or an attribute name is specified with a
colon. For example, rdf: specifies the RDF namespace; rdfs: specifies the RDFS
namespace; and owl: specifies the OWL namespace. Within an attribute value, a
namespace is specified with an XML entity. For example, &xsd; specifies the
XML Schema namespace. If no namespace is specified, the namespace is the current default namespace. In this case, the default namespace is obr, the namespace
of OBR. For example, OpeningFlagField is in the obr namespace.
At the highest level, the rule in Figure 13.9 says that an intersection of two
classes is a subclass of another class (lines 1, 2, 13, 14, 19, 20). This is expressed
using the owl:intersectionOf (lines 2, 13) and the rdfs:subClassOf (lines 14, 19)
properties. The intersection is used to represent the Boolean AND operator, and
subclass is used to present the logic IMPLIES or IF–THEN operator. The two
classes being intersected are the class of opening flag fields (line 3) and the
class of HDLC fields (lines 4–12). The former class is part of the ontology and
has a name (line 3). The latter class does not have a name. It is specified by two
relationships in the ontology: containedIn (line 5) and protocol (line 8). An HDLC
field is one that is contained in a frame whose protocol is of type HDLC. The
419
Chapter 13
owl:Restriction is used for constructing classes of instances that satisfy a constraint for
a particular property. The first restriction (line 4) specifies “a field that is contained in”
(lines 5, 6), and the second restriction (line 7) specifies “a frame whose protocol has
type HDLC” (lines 8, 9). The last restriction (line 15) specifies a value that must be
taken by the field (lines 16, 17). Putting all of these together, the rule can be expressed
as “IF something is an opening flag field AND is a field that is contained in a frame
whose protocol has type HDLC THEN it has value 7E. Alternatively, the same rule
could be represented in SWRL, as shown in Figure 13.10.
1 <ruleml:Imp>
2
<ruleml:body>
3
<swrlx:classAtom>
4
<owlx:Class owlx:name="#OpeningFlagField"/>
5
<ruleml:var>x</ruleml:var>
6
</swrlx:classAtom>
7
<swrlx:individualPropertyAtom swrlx:property="#containedIn">
8
<ruleml:var>x</ruleml:var>
9
<ruleml:var>y</ruleml:var>
10
</swrlx:individualPropertyAtom>
11
<swrlx:individualPropertyAtom swrlx:property="#protocol">
12
<ruleml:var>y</ruleml:var>
13
<ruleml:var>z</ruleml:var>
14
</swrlx:individualPropertyAtom>
15
<swrlx:classAtom>
16
<owlx:Class owlx:name="#HDLCProtocol"/>
17
<ruleml:var>z</ruleml:var>
18
</swrlx:classAtom>
19
</ruleml:body>
20
<ruleml:head>
21
<swrlx:datavaluedPropertyAtomsw rlx:property="#value">
22
<ruleml:var>x</ruleml:var>
23
<owlx:DataValue
24
owlx:datatype="&xsd;hexBinary">7E</owlx:DataValue>
25
</swrlx:datavaluedPropertyAtom>
26
</ruleml:head>
27 </ruleml:imp>
Figure 13.10: A SWRL implementation of the rule from Figure 13.9. This rule is used to
characterize the opening flag field of the DL layer frames.
At the highest level of the rule in Figure 13.10, there is an IF–THEN operator
(lines 1 and 27). An IF–THEN operator has two parts: the hypothesis or body (lines
2–19), and the conclusion or head (lines 20–26). Within either body or head, one
specifies a sequence of atoms. A class atom specifies an instance of a class. In this
case the class atoms specify that x is an instance of the class OpeningFlagField
(lines 3–6), and z is an instance of HDLCProtocol (lines 15–18). A property atom
specifies a triple. Individual property atoms are for properties whose values are
420
Roles of Ontologies in Cognitive Radios
individuals (lines 7–10 and 11–14). Data-valued property atoms are for properties
whose values are data values (lines 21–25). The second atom in the body specifies
the triple (x obr:containedIn y) (lines 7–10) and the third atom in the body specifies
the triple (y obr:protocol z) (lines 11–14). The head consists of a single atom,
which specifies the triple (x obr:value “7E”).
The full specification of all DL ontology rules in OWL is very large [17].
Rules dealing with multiple frames (such as a request and response to it) are especially complex because they are specifying functionality rather than just formats.
The DL ontology provides for the following capabilities:
1. Self-awareness of DL layer functionality: The ontology specifies the protocols.
More precisely, it specifies the frame structure and the functionality associated
with the various fields.
2. Dynamic interoperability at the DL layer: An OBR can infer the DL protocol
being used by another radio. A simple example of such a deduction was shown
by the list of properties 1 through 6 earlier in this section.
3. Command capability: A radio can remotely request the use of a different protocol or modify the features of an existing protocol.
13.5 Examples
Some examples of how two radios can exchange information about their communications characteristics and parameters and then adapt the communications protocol so that the quality of communication is maintained or improved demonstrate
the cognitive aspects described in the preceding sections.
13.5.1 Responding to Delays and Errors
Wireless transmission requires a robust and efficient communication protocol.
When the channel has been estimated and the estimation has been sent back to the
transmitter, then the transmission can be adapted according to the channel characteristics. The basic idea behind adaptive transmission is to maintain a constant
signal-to-noise ratio (SNR) level, (Eb /N0), by varying the transmission power
level, symbol transmission rate, constellation size, and coding rate/scheme, or any
combination of these parameters [7]. In experiments with protocol adaptation [7],
radios monitored their excess delay (multipath delay spread), rms delay (root
mean square delay spread) of the multipath structure of the channel, and the mean
square root error of the equalizer. Here the mean square root error of the equalizer
421
Chapter 13
represents the average distance between the equalized data (the input data of the
equalizer multiplied by the equalizer chips) and the output of the equalizer (the
estimated symbol).
OBRs are able to query each other by using a query expressed in an appropriate query language (e.g., OWL-QL). An example of such a query regarding the
rms delay, excess delay, and equalizer error is shown in Figures 13.11 and 13.12.
In this example, radio A first sends the following query to radio B (Figure 13.11).
<owl-ql:query ID=”query1”>
<owl-ql:queryPattern>
<rdf:RDF>
<obr:Multipath rdf:about="&obr;multipath">
<obr:rmsDelay rdf:resource="&var;x"/>
<obr:excessDelay rdf:resource="&var;y"/>
</obr:Multipath>
<obr:Equalizer rdf:about="&obr;equalizer">
<obr:equalizerError rdf:resource="&var;z"/>
</obr:Equalizer>
</rdf:RDF>
</owl-ql:queryPattern>
</owl-ql:query>
Figure 13.11: An OWL-QL query. The query to radio B from radio A is requesting information
concerning multipath parameters (rms delay and mean excess delay) as well as the meansquared decision error at the receiver B. The first line identifies the fragment as a query, and
creates the string identifier for it. The fourth line restricts attention to multipath parameters,
and the fifth and sixth lines request the particular variables rmsDelay (query variable x) and
excessDelay (query variable y) from that group. The eighth line refers to equalizer parameters,
and the ninth line requests the value of equalizerError (query variable z).
As was shown in Figure 13.2 for RDQL, a query consists of a series of triples in
which some of the slots contain variables. In OWL-QL, the query pattern is specified
by using RDF syntax. In RDF, the subject of a triple is specified by using rdf:about.
The first line of the query in Figure 13.2 specifies the triple (obr:multipath rdf:type
obr:Multipath) (i.e., obr:multipath is an instance of the obr:Multipath class). The second and third lines specify triples in which the subject is obr:multipath, the predicate
is the XML element tag, and the object is specified by using rdf:resource=. Thus the
third line specifies the triple (obr:multipath obr:rmsDelay var:x). In other words, the
variable x should be assigned to the rms delay of the multipath. The var namespace
contains the variables used by queries. Similarly, in the fifth line obr:equalizer is an
instance of the obr:Equalizer class, and the sixth line specifies the triple (obr:equalizer obr:equalizerError var:z).
422
Roles of Ontologies in Cognitive Radios
When radio B receives this query, it invokes its reasoner to derive an answer. The
query was formulated in the ontology that the two radios share, so the other radio
understands what obr:rmsDelay, obr:excessDelay, and obr:equalizer Error mean. For
instance, the ontology would include information on the units of measure for each
of the parameters. To answer the query, the radio first searches its annotation (or
markup) file to find or to infer facts that can match the query pattern. However, this
file contains only facts that have been explicitly asserted. It does not contain facts
that are embedded in the software. Consequently, the reflection mechanism is
invoked to extract facts that can be used to match the query pattern, either directly
or by means of inference. The answers to the query are then transmitted to radio
A. Such a response would look like the fragment in Figure 13.12.
<owl-ql:answerBundle about=”query1”>
<owl-ql:answer>
<owl-ql:binding-set>
<var:x>1.0078370372505556</var:x>
<var:y>1.062759005498691</var:y>
<var:z>0.025987243652343</var:z>
</owl-ql:binding-set>
</owl-ql:answer>
</owl-ql:answerBundle>
Figure 13.12: An OWL-QL response to radio A from radio B. The first line specifies the
fragment as a response, and identifies the corresponding query identifier (see Figure 13.11).
The fourth through sixth lines specify the values for query variables x, y, and z, which were
requested in the query query1, namely rmsDelay, excessDelay, and equalizerError.
Upon receiving this answer, radio A invokes its reasoner in order to decide how
to adjust its protocol so that the communications quality is improved. For instance,
the radio could do one of the following: select a different alphabet (e.g., 2-QAM,
4-QAM or 16-QAM), increase the equalizer length, or increase the transmit power.
13.5.2 Adaptation of Training Sequence Length
The following example considers the case of negotiating the length of the training
data according to the channel dynamics and noise level [8]. The goal is to have a
low packet overhead, when possible, and increase it in the case of high-noise or
high-delay spread. In these experiments, the transmitter used the DSSS spreading
type DSSS(2,7) for transmitting the header (i.e., each symbol is mapped to 127
chips), and a sequence of DSSS(2,5) symbols (31 chips for each symbol) for the
training data. The number of symbols in the training data was changed gradually
in response to the changing characteristics of the channel.
423
Chapter 13
In the example described in Wang et al. [8], negotiation of the length of the
training data was accomplished in six transmissions. First, node A sends data to
node B. This data comes to node B in an ontology-defined envelope (as is done in
Figure 13.11) as Data. Node B then checks performance. For this purpose, nodes
invoke their rules for performance checking. If performance is satisfactory, then
node B returns a Confirm message with content “Continue” and node A continues
to send data to B. If the performance is not satisfactory, a Confirm message with
content “CommandReq” (request a command from the other node) is returned.
Node A then generates a Command to change the communication protocol. The
command generation rules first send a Query from node A to node B requesting
the channel condition and the current protocol parameters (in a similar way as discussed in the previous example). When node B receives this query, it infers the
answer and sends the Answer back to node A. When node A receives the answer,
it generates the command. After the Command is generated, it is sent to node B,
and executed on node A, thus changing the protocol at node A. When node B
receives the Command, it executes the command, thus changing the protocol at
node B. A Confirm message with content “Continue” is then sent to node A. The
negotiation sequence involves five message types: data, confirmation, query,
answer, and command, as shown in the inheritance diagram in Figure 13.13.
These message types each have message contents, as described next.
Random Data
Content : String
Data
Confirm
Query
Answer
Command
Figure 13.13: Classification of message types.
To realize this kind of negotiation, the radios must not only share an ontology
but also have means for inferring ways of behaving in particular situations. In
pragmatic terms, this means that radios must have some domain knowledge of
radio communication. The knowledge may be represented in many different ways,
but as discussed in Section 13.2.1, the declarative representation is preferred to
the procedural. Consequently, this knowledge should be encoded in the form of
rules. For instance, the rule that is used in deriving a command concerning the
424
Roles of Ontologies in Cognitive Radios
number of symbols in the training sequence in response to this query might
look like this:
IF
●
the difference between the rmsDelay and the previous equalizer feedback coefficient vector length is larger than half of the previous equalizer
feedback coefficient vector length
THEN
●
construct a new equalizer with the length 3(integer(rmsDelay) 1) and
construct new training data of length 3 rmsDelay 20 MOD symbol
length
ELSE IF
the equalizer error is smaller than a predefined performance threshold
THEN
● decrease the length of the training data by a predefined value AND use
the old equalizer
●
ELSE
●
increase the length of the training data by a predefined value and use the
old equalizer
An informal textual representation of the rule is used here. The experiments
reported by Wang et al. [8] used a version of Prolog (Kernel Prolog) for this
purpose. In the future, the use of SWRL [16], is envisioned instead.
The experiments reported by Wang et al. [8] considered three different internode distances under different multipath conditions. It was shown that for the
worst condition, about 20 symbols were selected according to the negotiation
rules, whereas in other cases, only five symbols were needed to maintain a
desired level of quality of communication.
13.5.3 Data Link Layer Protocol Consistency and Selection
The DL ontology specifies the formats and functionality of the DL protocols. This
section presents a few examples of how the DL ontology might be used. The first
example illustrates how the formal specification of protocols can uncover inconsistencies. This kind of reasoning would normally be performed offline. The second example shows a simple example of how the ontology can be used for
interoperability.
425
Chapter 13
The first example uses the following two rules (2 and 5) from the DL ontology
list presented in Section 13.4.2:
●
●
If a frame belongs to the SDLC protocol, then the address field has 8 bits.
If a frame belongs to the HDLC protocol and the mode is ARM, then the
address field has 0 bits.
Because the SDLC protocol is a subclass of the HDLC protocol, the two rules
imply that an ARM frame belonging to the SDLC protocol has an address field
with both 8 and 0 bits. It follows that the SDLC protocol cannot have ARM
frames.
We can address this inconsistency in several ways. We could remove the subclass relationship between the SDLC protocol and the HDLC protocol classes
from the ontology. This has the disadvantage that the SDLC protocol would now
have to be completely specified from scratch rather than being derived from a
more general class. Alternatively, we could modify the first of these two rules to
recognize that the address field for the SDLC protocol will have only 8 bits for
modes other than ARM.
The second example concerns the issue of interoperability. Suppose that a
radio receives a stream of packets from another radio with which it has not previously communicated. Can the receiving radio deduce the protocol being used by
the transmitter? In the case of HDLC versus WiFi, the following two rules (1 and
6 from the list in Section 13.4.2) allow one to disambiguate the packets:
●
●
If a frame belongs to the HDLC protocol, then its opening flag field has value
0 7E (i.e., the bit sequence 01111110).
If a frame belongs to the WiFi protocol then the first two bits of the control field
are zeroes. This subfield represents the version number.
From these two rules, we can deduce that the first two bits of the packet will
distinguish these two protocols. If the first two bits are 01, then it is an HDLC
frame. If the first two bits are 00, then it is a WiFi frame. If the bits have some
other value, then the frame belongs to neither protocol. This example is somewhat
artificial, and we could argue that it is easy to write a small procedure that could
make this same determination. However, this is only one example among many
possible inferences. The software would quickly become extremely unwieldy if
we needed a separate procedure for every possible deduction. Indeed, we would
426
Roles of Ontologies in Cognitive Radios
need infinitely many procedures because we can make infinitely many deductions,
and every time a new feature or protocol variation was added, we would need to
revise all of the procedures.
13.6 Open Research Issues
Even though the concept of OBR can be implemented as discussed in this chapter, a
number of open issues still need to be investigated. Additionally, some of the engineering issues need to be addressed before this concept finds its way into real radio
applications. Several of the most important open issues are discussed in this section.
13.6.1 Ontology Development and Consensus
As discussed in this chapter, ontologies can bring many advantages to software
radio communication. However, to use an ontology, one must first develop it.
Building high-quality ontologies is a substantial task. It is further complicated by
the diversity of web-based ontology languages.
The first step in an ontology development project is to agree upon the purpose
and motivation for embarking on this activity. This step includes determining the
following:
1.
2.
3.
4.
5.
Why the ontology is being developed.
What will be covered by the ontology.
Who will be using the ontology.
When and for how long the ontology will be used.
How the ontology is intended to be used.
Once there is a clear understanding of the purpose of the ontology, four major
activities must be undertaken: (1) choosing an ontology language, (2) obtaining a
development tool, (3) acquiring domain knowledge, and (4) reusing existing
ontologies. None of these is especially easy, and it is important to reach consensus
because the ontology forms the basis for communication within the community.
Generally speaking, the larger the community, the more difficult it is to reach consensus. One can mitigate this difficulty to some degree by establishing a precise
statement of the purpose of the ontology before starting the project. However, one
cannot entirely eliminate it.
Ultimately, the success of ontologies for software and cognitive radio communication will depend not only on the quality of the ontology design but also on the
427
Chapter 13
extent to which the community accepts the ontology. The task is a challenging one,
but experience in other communities suggests that it is achievable, nonetheless.
13.6.2 Ontology Mapping
At every level of the communication protocol stack are many protocols currently
in use. Each protocol has its own terminology and therefore its own ontology.
Generally speaking, the higher in the stack, the more diverse the ontologies. To
achieve interoperability between protocols and other applications, it is necessary
to have mappings between the ontologies. Such mappings can be specified by subject matter experts, but when the ontologies use the same concepts, it should be
possible to automate the process of transforming from one ontology to the other.
This problem has been studied for many years. Most of the work has been devoted
to mapping relational database schemas, but there has also been some recent work
on this problem for XML DTDs and even for the more sophisticated ontologies of
the Semantic Web.
A survey of relational schema integration tools is presented by Rahm and
Bernstein [18]. When the data from a variety of relational database sources are
transformed to a single target database, the process is called data warehousing.
Many data warehousing companies now also support XML. If a query using one
vocabulary is rewritten so as to retrieve data from various sources, each of which
uses its own vocabulary, it is called virtual data integration [19].
Ontology mapping depends on identifying semantically corresponding elements. This is a difficult problem to solve because different sources may use very
different structural and naming conventions for the same entities. In addition, the
same name can be used for elements having totally different meanings, such as
different units, precision, resolution, measurement protocol, and so on. It is usually necessary to annotate an ontology with auxiliary information to assist one in
determining the meaning of elements, but ontology mapping is difficult to automate even with this additional information. Furthermore, a single element in one
ontology may correspond to multiple elements in another. In general, the correspondence between elements is many-to-many: many elements correspond to
many elements.
Tools for automating ontology mapping have been proposed, and some
research prototypes exist. However, these tools mainly help to discover simple
one-to-one matches, and they do not consider the meaning of the data or how
the transformation will be used. Using such a tool requires significant manual
effort to correct wrong matches and add missing matches. One of the few tools
428
Roles of Ontologies in Cognitive Radios
that can handle many-to-many mappings was developed by Goguen and his
students [20].
In practice, ontology mapping is done manually by domain experts and is very
time-consuming when there are many data sources, or when ontologies are large
or complex. Automated ontology-mapping systems are designed to reduce manual
effort. However, such systems require a substantial amount of time to prepare
input to the system as well as to guide the matching process. This amount of time
may easily swamp time saved by using the system. Unfortunately, existing
ontology-mapping prototypes focus on measuring accuracy and completeness
rather than on whether they provide a net gain. In practice, the best that one can
hope for from current systems is that they can help to record and to manage the
ontology matches that have been detected, by whatever means.
Many ontology-mapping projects exist, and some have developed prototypes.
Examples are PROMPT [21], from the Stanford Medical Informatics Laboratory,
and the Semantic Knowledge Articulation Tool (SKAT) [22], also from Stanford.
None of these prototypes are available for public use, however, either via open
source software or commercial software. But even if the tools were available, they
are only prototypes, not developed to deal with the issues that are specific to radio
communication, such as reflection and high-performance reasoning. Consequently,
ontology mapping remains a challenging field for development.
13.6.3 Learning
Ontology learning is an emerging field aimed at automated assistance in the
construction of ontologies and semantic page annotation through the use of
machine learning techniques [23]. The relevance of this field is tied to the prohibitively large amount of manual ontological construction required to represent
the vast amount of information in communication systems. The implementation of a fixed ontological system presumes that the knowledge engineer
anticipated the extent of the language involved in queries, responses, and reasoning. Although this goal could be attained at some time instant, the continual evolution of radio standards and implementations mandates a requisite evolution in
the ontology. For example, WiFi was not a recognized radio standard 15 years
ago, IEEE802.11g was not available prior to 2000, and the pre-release version of
802.11n was not available until 2004. Additionally, new variants and derivatives
(subclasses) of the HDLC protocol class will continue to appear for some time.
Without automated ontological learning, a cognitive radio that is expected to
communicate using these emerging standards would require continuous updating
429
Chapter 13
of its ontological database system. Ontology learning is one mechanism to automate this evolution.
13.6.4 Efficiency of Reasoning
The implementation of the concept of OBR discussed in this chapter relies heavily
on OWL, a language for representing ontologies. Another component of the OBR
architecture is a generic reasoning engine. Even though a generic reasoning mechanism satisfies the requirement of being able to respond to surprise by a cognitive
agent, it also poses very high demand on the reasoning engine. In general, logic
reasoning is undecidable. In simple terms, this means that an inference engine
might never terminate its process of answering a specific query.
With this concern in mind, the designers of OWL provided the language in
three levels: OWL Full, OWL-DL, and OWL Lite, as discussed in section
Ontology Languages. The three levels differ in both expressive power and computational complexity. OWL Lite is the simplest of the three, with the least expressive power of the three OWL languages. In spite of its relative simplicity, the
computational complexity of OWL Lite is very high; it is in the EXPTIME complexity class, meaning that in the worst case a deduction will require time equal
to O(2p(n)), where p(n) is a polynomial in the size n of the ontology. OWL-DL is
more expressive than OWL Lite, but its complexity is even worse; it is in the
NEXPTIME complexity class. Although the preside complexity of this class is
not known, it is at least as high as EXPTIME, and it may be much higher still,
possibly an exponential of an exponential of a polynomial (i.e., O(2f(n)), where
f(n) 2p(n). Furthermore, there is no bound on how much memory may be
required. OWL Full is the most expressive of the three OWL languages, but its
complexity is the worst possible: it is undecidable. Undecidability means that
in the worst case there is no bound on the amount of time or space that may be
needed to perform a deduction. In practice, even OWL Full is not sufficiently
expressive, and it is necessary to use rules to express some of the axioms of the
ontology. Unfortunately, if rules are added to any of the three languages, including
OWL Lite, then the computational complexity of deduction is undecidable.
In order to make systems based on such generic reasoners scalable, the issues
of complexity of reasoning must be resolved so that reasoning conclusions are
derived within the constraints of the radio domain. Various approaches to such a
problem are known in the literature. In general, such approaches are based on a
trade-off between the quality of the solution and the computation time. For instance,
for algorithms known as any time algorithms, the problem-solving activity can be
430
Roles of Ontologies in Cognitive Radios
stopped at any time, and the result will be the best one that has been found up to
that time. Another approach is to combine logic reasoning with an uncertainty
handling mechanism, like Bayesian networks, fuzzy logic, and others.
13.7 Summary
This chapter presented an ontology-based approach to address some of the
requirements of cognitive radio. Throughout, this chapter illustrated how we can
capture knowledge about the domain of radio communication using ontologies.
All of the examples used a purely declarative representation of domain knowledge. The objective was to show that such a knowledge representation can be used
by generic reasoning tools both to formulate queries and to infer answers to such
queries. Such a query-answering mechanism is generic and thus satisfies the
requirement of being capable of “responding to surprise,” one of the key features
of a cognitive agent.
Another feature of a cognitive agent is the ability to collect and use information about the environment. The example in Section 13.5.1 showed how radios
can request multipath information and then use it to modify the transmission protocol. Moreover, such radios are situation-aware because they know not only
transmission conditions at other nodes, but also how such conditions relate to their
own transmission protocol parameters. Section on Querying also discussed how
self-awareness can be implemented so that radios can know the values of their
internal variables and how to change those values on demand. For instance, a
radio can know its rms delay, excess delay, or equalizer error. Another example of
this capability, shown in Section 13.5.2, is the knowledge of such values as the
BER and the multipath structure.
This chapter also showed examples of querying of radios by other radios.
Although it did not show any queries by end users, such queries can be implemented in the same manner as queries by other radios. This chapter has also
shown that self-aware radios can execute commands. For instance, if two radios
agree to use a specific length of the training sequence, they first negotiate a contract, and then both update their internal structures. In general, any information
expressible in terms of the ontology can be retrieved and modified.
References
[1] R. Brachman, “A DARPA Information Processing Technology Renaissance: Developing Cognitive Systems.” http://www.darpa.mil/ipto/presentations/brachman.ppt
431
Chapter 13
[2] J. Boyd, A Discourse on Winning and Losing, Technical Report, Maxwell AFB, 1987.
[3] K.M. Passino and P.J. Antsaklis (Eds.), Introduction to Intelligent and Autonomous
Control, Kluwer Academic Publishers, Boston, Dodrech, /London, 1992.
[4] M. Endsley and D. Garland, Situation Awareness, Analysis and Measurement,
Lawrence Erlbaum Associates, Publishers, Mahway, New Jersey, 2000.
[5] XG Working Group. The XG Vision. Request For Comments, Version 2.0.
[6] N. Guarino, “Formal Ontology in Information Systems,” in Proceedings of Formal
Ontology in Information Systems, IOS Press, Amsterdam, 1998, pp. 3–15.
[7] J. Wang, D. Brady, K. Baclawski, M. Kokar and L. Lechowicz, “The Use of
Ontologies for the Self-Awareness of the Communication Nodes,” in Proceedings
of the Software Defined Radio Technical Conference SDR ’03, Orlando, Florida,
USA, 2003.
[8] J. Wang, M.M. Kokar, K. Baclawski and D. Brady, “Achieving Self-Awareness of
SDR Nodes Through Ontology-Based Reasoning and Reflection,” in Proceedings
of the Software Defined Radio Technical Conference SDR ’04, Phoenix, Arizona,
November 2004.
[9] World Wide Web Consortium, “eXtensible Markup Language,” www.w3.org/
XML/, 2001.
[10] TopicMaps.org, The XTM website, topicmaps.org, 2000.
[11] O. Lassila and R. Swick, “Resource Description Framework (RDF) Model and
Syntax Specification,” www.w3.org/TR/REC-rdf-syntax, February 1999.
[12] F. van Harmelen, J. Hendler, I. Horrocks, D. McGuinness, P. Patel-Schneider and
L. Stein, “OWL Web Ontology Language Reference,” www.w3.org/TR/owl-ref/
M. Dean and G. Schreiber (Eds.), March 2003 .
[13] F. Baader, D. Nardi, D. McGuinness, P. Patel-Schneider and D. Calvanese. The
Description Logic Handbook: Theory, Implementation and Applications,
Cambridge University Press, Cambridge, England, 2003.
[14] A. Seaborne. “RDQL—A Query Language for RDF,” www.w3.org/Submission/
RDQL, January 2004.
[15] R. Fikes, P. Hayes and I. Horrocks, “OWL-QL: A Language for Deductive Query
Answering on the Semantic Web,” Journal of Web Semantics, Vol. 2, No. 2, 2005.
[16] I. Horrocks, P. Patel-Schneider, H. Boley, S. Tabet, B. Grosof and M. Dean,
“SWRL: A Semantic Web Rule Language Combining OWL and RuleML,”
http://www.daml.org/2003/11/swrl, 2003.
[17] D. Brady, K. Baclawski and M. Kokar, “Achieving Dynamic Interoperability
of Communication at the DataLink Layer through Ontology Based Reasoning,”
in Proceedings of SDR Forum Tech Conference SDR05, Anaheim, CA,
November 2005.
[18] E. Rahm and P. Bernstein, “On Matching Schemas Automatically,” VLDB Journal,
Vol. 10, No. 4, 334–350, 2001.
432
Roles of Ontologies in Cognitive Radios
[19] D. Embley, et al., “Multifaceted Exploitation of Metadata for Attribute Match
Discovery in Information Integration,” In Proceedings of the International
Workshop on Information Integration on the Web (WIIW'01), pages 110--117, Rio
de Janeiro, Brazil, April 2001.
[20] Y. Nam, J. Goguen and G. Wang, “ A Metadata Integration Assistant Generator for
Heterogeneous Distributed Databases,” in Proceedings of International Conference
Ontologies, Databases, and Applications of Semantics for Large Scale Information
Systems, Lecture Notes in Computer Science 2519, Springer-Verlag, Berlin,
Heidelberg, New York, 2002, 1332–1344.
[21] N. Noy and M. Musen, PROMPT: Algorithm and Tool for Automated Intelligence,
AAAI-2000, Austin TX, 2000.
[22] P. Mitra, G. Weiderhold and J. Jannink, “Semi-automatic Integration of Knowledge
Sources,” in Proceedings of 2nd International Conference on Information Fusion,
Sunnyvale, CA, 1999.
[23] B. Omelayenko, “Learning of Ontologies for the Web: The Analysis of Existent
Approaches,” in Proceedings of International Conference on Database Theory,
London, UK, January, 2001.
433
This page intentionally left blank
CHAPTER 14
Cognitive Radio Architecture
Joseph Mitola III
The MITRE Corporation
Tampa, FL, USA
14.1 Introduction
Architecture is a comprehensive, consistent set of design rules by which a
specified set of components achieves a specified set of functions in products
and services that evolve through multiple design points over time [1]. This
section introduces the fundamental design rules by which software-defined radio
(SDR), sensors, perception, and automated machine learning (AML) may be
integrated to create aware, adaptive, and cognitive radios (AACRs). These
SDRs will have better quality of information (QoI) through capabilities to observe
(sense, perceive), orient, plan, decide, act, and learn (the so-called OOPDAL
loop) in radio frequency (RF) and in the user domains. By performing this
integration, we will transition from merely adaptive to a demonstrably cognitive
radio (CR).
This section develops five complementary perspectives of CR architecture
(CRA), called CRA I through CRA V. The CRA I perspective defines six functional components, black boxes to which are ascribed a first-level decomposition
of AACR functions and among which important interfaces are defined. One of
these boxes is SDR, a proper subset of AACR. One of these boxes performs
Note: The MITRE Corporation is provided for identification purposes only and should not be interpreted as the endorsement of the material by the MITRE Corporation or any of its sponsors. Adapted
from Cognitive Radio Architecture: The Engineering Foundations of Radio XML, Wiley, 2006.
435
Chapter 14
cognition via the <Self/>,1 a self-referential subsystem that strictly embodies finite
computing (e.g., no While or Until loops) avoiding the Gödel-Turing paradox.2
The CRA II perspective examines the flow of inference through a cognition
cycle that arranges the core capabilities of ideal CR (iCR) in temporal sequence for
a logical flow and circadian rhythm for the CRA. The CRA III perspective examines the related levels of abstraction for AACR to sense elementary sensory stimuli
and to perceive QoI relevant aspects of a <Scene/> consisting of the <User/> in an
<Environment/> that includes <RF/>.3 The CRA IV perspective examines the mathematical structure of this architecture, identifying mappings among topological
spaces represented and manipulated to preserve set-theoretic properties. Finally,
the CRA V perspective reviews SDR architecture, sketching an evolutionary path
from the Software Communications Architecture/Software Radio Architecture
(SCA/SRA) to the CRA. The CRA is expressed in Radio eXtensible Markup
Language (RXML). The CRA is introduced in this chapter and developed in [12].
14.2 CRA I: Functions, Components, and Design Rules
The functions of AACR exceed those of SDR. Reformulating the AACR <Self/>
as a peer of its own <User/> establishes the need for added functions by which
the <Self/> accurately perceives the local scene including the <User/> and
autonomously learns to tailor the information services to the specific <User/> in the
current RF and physical <Scene/>.
14.2.1 AACR Functional Component Architecture
The SDR components and the related cognitive components of iCR appear in
Figure 14.1. The cognition components describe the SDR in RXML so that the
<Self/> can know that it is a radio and that its goal is to achieve high QoI tailored
to its own users. RXML intelligence includes a priori radio background and user
stereotypes as well as knowledge of RF and space–time <Scenes/> perceived and
experienced. This includes both structured reasoning with iCR peers and cognitive
1
Note that <Self/> is how the CR will refer to itself, so “<Self/>” can also be read as “the radio.”
This refers to the classic Turing machine, which can fall into unbounded analysis when attempting to reason about itself. One solution to this problem is a timer that kills any task that isn’t completed within a specified time.
3
The use of Semantic Web technology (e.g., referring to the <User/> in an <Environment/>) is
addressed by Mahonen [2].
2
436
Cognitive Radio Architecture
wireless networks (CWNs), and ad hoc reasoning with users, all the while learning
from experience.
CR1
Radio Knowledge
RXML
Representation
Language
Radio Knowledge
RKRL
Structured and Ad Hoc Reasoning,
Learning from Experience
Cognition
User Knowledge
Equalizer
Antenna
RF
Modem
INFOSEC
XML
User Interface
SWR
Software Modules .....
RAM
Baseband
Back End Control
Baseband Modem
Equalizer
Algorithm
Software
SDR
Hardware
Antenna
RF
Modem
INFOSEC
Baseband
User Interface
Figure 14.1: The CRA augments SDR with computational intelligence and learning capacity
(© Dr. Joseph Mitola III, used with permission).
The detailed allocation of functions to components with interfaces among the
components requires closer consideration of the SDR component as the foundation of CRA.
SDR Components
SDRs include a hardware platform with RF access and computational resources,
plus at least one software-defined personality. The SDR Forum has defined its
SCA [3] and the Object Management Group (OMG) has defined its SRA [4].
These are similar fine-grained architecture constructs enabling reduced-cost wireless connectivity with next-generation plug-and-play. These SDR architectures
are defined in Unified Modeling Language (UML) object models [5], Common
Object Request Broker Architecture (CORBA) Interface Design Language (IDL)
[7], and XML descriptions of the UML models. The SDR Forum and OMG
standards describe the technical details of SDR both for radio engineering and
for an initial level of wireless air interface (“waveform”) plug-and-play. The
SCA/SRA was sketched in 1996 at the first US Department of Defense (DoD)
inspired modular multifunctional information transfer system (MMITS) Forum,
437
Chapter 14
was developed by the DoD in the 1990s and the architecture is now in use by the
US military [7]. This architecture emphasizes plug-and-play wireless personalities
on computationally capable mobile nodes where network connectivity is often
intermittent at best.
The commercial wireless community [8], in contrast, led by cell phone giants
Motorola, Ericsson, and Nokia, envisions a much simpler architecture for mobile
wireless devices, consisting of two application programming interfaces (APIs)—one
for the service provider and another for the network operator. Those users define a
knowledge plane in the future intelligent wireless networks that is not dissimilar
from a distributed CWN. That community promotes the business model of the
user → service provider → network operator → large manufacturer → device, in
which the user buys mobile devices consistent with services from a service provider,
and the technical emphasis is on intelligence in the network. This perspective no
doubt will yield computationally intelligent networks in the near- to mid-term.
The CRA developed in this text, however, envisions the computational intelligence to create ad hoc and flexible networks with the intelligence in the mobile
device. This technical perspective enables the business model of user → device →
heterogeneous networks, typical of the Internet model in which the user buys a
device (e.g., a wireless laptop) that can connect to the Internet via any available
Internet service provider (ISP). The CRA builds on both the SCA/SRA and the
commercial API model, but integrates Semantic Web intelligence in RXML for
more of an Internet business model. This chapter describes how SDR, AACR, and
iCR form a continuum facilitated by RXML.
AACR Node Functional Components
A simple CRA includes the functional components shown in Figure 14.2. A functional component is a black box to which functions have been allocated, but for
which implementation is not specified. Thus, while the applications component
is likely to be primarily software, the nature of those software components is yet
to be determined. User interface functions, however, may include optimized hardware (e.g., for computing video flow vectors in real time to assist scene perception). At the level of abstraction of this figure, the components are functional,
not physical.
These functional components are as follows:
1. The user sensory perception (SP), which includes haptic, acoustic, and video
sensing and perception functions.
438
Cognitive Radio Architecture
User Interface
Functions
Applications
User
Radio Networks
Environment sensor
functions
Environment
Effector
Functions
SDR
Functions
Other Networks
Cognition
Functions
CR
Figure 14.2: Minimal AACR node architecture (© Dr. Joseph Mitola III, used with
permission).
2. The local environment sensors (location, temperature, accelerometer,
compass, etc.).
3. The system applications (sys apps) media-independent services such as playing
a network game.
4. The SDR functions which include RF sensing and SDR applications.
5. The cognition functions (symbol grounding for system control, planning, and
learning).
6. The local effector functions (speech synthesis, text, graphics, and multimedia
displays).
These functional components are embodied on an iCR platform, a hardware
realization of the six functions. To support the capabilities described in the prior
chapters, these components go beyond SDR in critical ways. First, the user interface goes well beyond buttons and displays. The traditional user interface has
been partitioned into a substantial user sensory subsystem and a set of local effectors. The user sensory interface includes buttons (the haptic interface) and microphones (the audio interface) to include acoustic sensing that is directional, capable
of handling multiple speakers simultaneously, and able to include full motion
video with visual scene perception. In addition, the audio subsystem does not just
encode audio for (possible) transmission; it also parses and interprets the audio
from designated speakers, such as the <User/>, for a high-performance spoken
natural language (NL) interface. Similarly, the text subsystem parses and interprets the language to track the user’s information states, detecting plans and
potential communications and information needs unobtrusively as the user
439
Chapter 14
conducts normal activities. The local effectors synthesize speech along with traditional text, graphics, and multimedia displays.
Sys apps are those information services that define value for the user.
Historically, voice communications with a phone book, text messaging, and the
exchange of images or video clips comprised the core value proposition of sys
apps for SDR. These applications were generally integral to the SDR application,
such as data services via general packet radio service (GPRS), which is really a
wireless SDR personality more than an information service. AACR sys apps
break the service out of the SDR waveform, so that the user need not be limited
by details of wireless connectivity unless that is of particular interest. Should the
user care whether he or she plays the distributed video game via 802.11 or Bluetooth
over the last 3 m? Probably not. The typical user might care if the AACR wants to
switch to third generation (3G) at $5 per minute, but a particularly affluent user
might not care and would leave all that up to the AACR.
The cognition component provides all the cognition functions—from the
semantic grounding of entities in the perception system to the control of the overall system through planning and initiating actions—learning user preferences and
RF situations in the process.
Each of these subsystems contains its own processing, local memory, integral
power conversion, built-in test (BIT), and related technical features.
The Ontological <Self/>
AACR consists of six functional components: user SP, environment, effectors,
SDR, sys apps, and cognition. Those components of the <Self/> enable external
communications and internal reasoning about the <Self/> by using the RXML
syntax. Given the top-level outline of these functional components along with the
requirement that they be embodied in physical hardware and software (the “platform”), the six functional components are defined ontologically in the equation in
Figure 14.3. In part, this equation states that the hardware–software platform and
<iCR-Platform/>
<Functional-Components>
<User SP/><Environment/><Effectors/><SDR/><Sys Apps/><Cognition/>
</Functional-Components>
</Self>
Figure 14.3: Components of the AACR <Self/>. The AACR <Self/> is defined to be an iCR
platform, consisting of six functional components using the RXML syntax.
440
Cognitive Radio Architecture
the functional components of the AACR are independent. Platform-independent
computer languages such as Java are well understood. This ontological perspective envisions platform independence as an architecture design principle for
AACR. In other words, the burden is on the (software) functional components to
adapt to whatever RF–hardware–operating system platform might be available.
14.2.2 Design Rules Include Functional Component Interfaces
The six functional components (see Tables 14.1(a) and 14.1(b)) imply
associated functional interfaces. In architecture, design rules may include a list of
the quantities and types of components as well as the interfaces among those
components. This section addresses the interfaces among the functional
components.
The AACR N-squared diagram of Table 14.1(a) characterizes AACR interfaces. These constitute an initial set of AACR APIs. In some ways, these APIs
augment the established SDR APIs. This is entirely new and much needed in
order for basic AACRs to accommodate even the basic ideas of the Defense
Advanced Research Projects Agency (DARPA) NeXt-Generation (XG) radio
communications program.
In other ways, these APIs supersede the existing SDR APIs. In particular, the
SDR user interface becomes the user sensory and effector API. User sensory APIs
include acoustics, voice, and video, and the effector APIs include speech synthesis
Table 14.1(a): AACR N-squared diagram. This matrix characterizes internal interfaces
between functional processes. Interface notes 1–36 are explained in Table 14.1(b).
From\to
User SP
Environment Sys apps
User SP
1
2
3
4
5 PECb
6 SC
7
8
9
10
11 PECb
12
Environment
Sys apps
SDR
Cognition
Effectors
13 PAa
14 SAa
15 SCMa
16 PDa
17 PCa,b
18a
SDR
Cognition
Effectors
19
20
21 SDa
22 SD
23 PAEb
24
25 PAb
26 PAb
27 PDCa,b
28 PCb
29 SCb
30 PCDb
31
32
33 PEMa
34 SD
35 PEb
36
P: primary; A: afferent; E: efferent; C: control; M: multimedia; D: data; S: secondary; others not
designated P or S are ancillary.
a
Information services API; bCAPI.
441
Table 14.1(b): Explanations of interface notes for functional processes shown in Table 14.1(a).
Note
number
Process interface
Explanation
1
User SP–
User SP
2
Environment–
User SP
Sys apps–
User SP
SDR–User SP
Cross-media correlation interfaces (video-acoustic, haptic-speech, etc.) to limit
search and reduce uncertainty (e.g., if video indicates user is not talking, acoustics
may be ignored or processed less aggressively for command inputs than if user is
speaking).
Environment sensors parameterize user sensor-perception. Temperature below
freezing may limit video.
Sys apps may focus scene perception by identifying entities, range, expected sounds
for video, audio, and spatial perception processing.
SDR applications may provide expectations of user input to the perception system
to improve probability of detection and correct classification of perceived inputs.
This is the primary control efferent path from cognition to the control of the user
SP subsystem, controlling speech recognition, acoustic signal processing, video
processing, and related SP. Plans from cognition may set expectations for user scene
perception, improving perception.
Effectors may supply a replica of the effect to user perception so that self-generated
effects (e.g., synthesized speech) may be accurately attributed to the <Self/>,
validated as having been expressed, and/or canceled from the scene perception to
limit search.
Perception of rain, buildings, indoor/outdoor can set GPS integration parameters.
3
442
4
5
Cognition–
User SP
6
Effectors–
User SP
7
User SP–
Environment
Environment–
Environment
8
Environment sensors would consist of location sensing such as GPS or GLONASS;
ambient temperature; light level to detect inside versus outside locations; possibly
9
10
11
12
443
13
14
Sys apps–
Environment
SDR–
Environment
Cognition–
Environment
(primary
control path)
Effectors–
Environment
UserSP–Sys
apps
Environment–
Sys apps
smell sensors to detect spoiled food, fire, etc. There seems to be little benefit in
enabling interfaces among these elements directly.
Data from the sys apps to environment sensors would also be minimal.
Data from the SDR personalities to the environment sensors would be minimal.
Data from the cognition system to the environment sensors controls those sensors,
turning them on and off, setting control parameters, and establishing internal paths
from the environment sensors.
Data from effectors directly to environment sensors would be minimal.
Data from the user SP system to sys apps is a primary afferent path for multimedia
streams and entity states that effect information services implemented as sys apps.
Speech, images, and video to be transmitted move along this path for delivery by the
relevant sys apps or information service to the relevant wired or SDR communications path. Sys apps overcomes the limitations of individual paths by maintaining
continuity of conversations, data integrity, and application coherence (e.g., for
multimedia games). Whereas the cognition function sets up, tears down, and
orchestrates the sys apps, the primary API between the user scene and the information service consists of this interface and its companions—the environment afferent
path, the effector efferent path, and the SDR afferent and efferent paths.
Data on this path assists sys apps in providing location awareness to services.
(Continued)
Table 14.1(b): Explanations of interface notes for functional processes shown in Table 14.1(a). (Continued)
Note
number
444
Process interface
Explanation
15
Sys apps–
Sys apps
16
SDR–Sys apps
17
Cognition–
Sys apps
Effectors–
Sys apps
Different information services interoperate by passing control information through
the cognition interfaces and by passing domain multimedia flows through this interface. The cognition system sets up and tears down these interfaces.
This is the primary afferent path from external communications to the AACR. It
includes control and multimedia information flows for all the information services.
Following the SDR Forum’s SCA, this path embraces wired as well as wireless
interfaces.
Through this path, the AACR <Self/> exerts control over the information services
provided to the <User/>.
Effectors may provide incidental feedback to information services through this afferent path, but the use of this path is deprecated. Information services are supposed to
control and obtain feedback through the mediation of the cognition subsystem.
Although the SP system may send data directly to the SDR subsystem (e.g., to satisfy security rules that user biometrics be provided directly to the wireless security
subsystem), the use of this path is deprecated. Perception subsystem information is
supposed to be interpreted by the cognition system so that accurate information, not
raw data, can be conveyed to other subsystems.
Environment sensors such as GPS historically have accessed SDR waveforms
directly (e.g., providing timing data for air interface signal generation). The cognition system may establish such paths in cases where cognition provides little or no
value added, such as providing a precise timing reference from GPS to an SDR
waveform. The use of this path is deprecated because all of the environment sensors,
18
19
User SP–SDR
20
Environment–
SDR
21
22
23
445
24
25
26
including GPS, are unreliable. Cognition has the capability to “de-glitch” GPS (e.g.,
recognize from video that the <Self/> is in an urban canyon and therefore not allow
GPS to report directly, but report to the GPS subscribers, on behalf of GPS, location
estimates based perhaps on landmark correlation, dead reckoning, etc.).
Sys apps–SDR This is the primary efferent path from information services to SDR through the
services API.
SDR–SDR
The linking of different wireless services directly to each other is deprecated. If an
incoming voice service needs to be connected to an outgoing voice service, there
should be a bridging service in sys apps through which the SDR waveforms communicate with each other. That service should be set up and taken down by the cognition system.
Cognition–SDR This is the primary control interface, replacing the control interface of the SDR SCA
and the OMG SRA.
Effectors–SDR Effectors such as speech synthesis and displays should not need to provide state
information directly to SDR waveforms, but if needed, the cognition function should
set up and tear down these interfaces.
User SP–
This is the primary afferent flow for the results from acoustics, speech, images,
Cognition
video, video flow, and other sensor-perception subsystems. The primary results
passed across this interface should be the specific states of <Entities/> in the scene,
which would include scene characteristics such as the recognition of landmarks,
known vehicles, furniture, and the like. In other words, this is the interface by which
the presence of <Entities/> in the local scene is established and their characteristics
are made known to the cognition system.
Environment– This is the primary afferent flow for environment sensors.
Cognition
(Continued)
Table 14.1(b): Explanations of interface notes for functional processes shown in Table 14.1(a). (Continued)
Note
number
27
28
446
29
30
31
Process interface
Sys apps–
Cognition
Explanation
This is the interface through which information services request services and receive
support from the AACR platform. This is also the control interface by which cognition sets up, monitors, and tears down information services.
SDR–Cognition This is the primary afferent interface by which the state of waveforms, including a
distinguished RF-sensor waveform, is made known to the cognition system. The
cognition system can establish primary and backup waveforms for information services, enabling the services to select paths in real time for low-latency services. Those
paths are set up and monitored for quality and validity (e.g., obeying XG rules) by
the cognition system, however.
The cognition system as defined in this six-component architecture entails (1) orientCognition–
Cognition
ing to information from <RF/> sensors in the SDR subsystem and from scene
sensors in the user SP and environment sensors; (2) planning; (3) making decisions;
and (4) initiating actions, including the control over all of the cognition resources of
the <Self/>. The <User/> may directly control any of the elements of the systems via
paths through the cognition system that enable it to monitor what the user is doing in
order to learn from a user’s direct actions, such as manually tuning in the user’s
favorite radio station when the <Self/> either failed to do so properly or was not asked.
Effectors–
This is the primary afferent flow for status information from the effector subsystem,
Cognition
including speech synthesis, displays, and the like.
User SP–
In general, the user SP system should not interface directly to the effectors, but
Effectors
should be routed through the cognition system for observation.
32
33
Environment–
Effectors
Sys apps–
Effectors
447
34
SDR–Effectors
35
Cognition–
Effectors
36
Effectors–
Effectors
The environment system should not interface directly to the effectors. This path is
deprecated.
sys apps may display streams, generate speech, and otherwise directly control any
effectors once the paths and constraints have been established by the cognition
subsystem.
This path may be used if the cognition system establishes a path, such as from an
SDR’s voice track to a speaker. Generally, however, the SDR should provide streams
to the information services of the sys apps. This path may be necessary for legacy
compatibility during the migration from SDR through AACR to iCR, but it is
deprecated.
This is the primary efferent path for the control of effectors. Information services
provide the streams to the effectors, but cognition sets them up, establishes paths, and
monitors the information flows for support to the user’s <Need/> or intent.
These paths are deprecated, but may be needed for legacy compatibility.
Chapter 14
to give the AACR <Self/> its own voice. In addition, wireless applications are
growing rapidly. Voice and short-message service provide an ability to exchange
images and video clips with ontological tags among wireless users. The distinctions between cell phone, personal digital assistant (PDA), and game box continue
to disappear.
These interface changes enable the AACR to sense the situation represented in
the environment, to interact with the user, and to access radio networks on behalf
of the user in a situation-aware way.
The above information flows aggregated into an initial set of AACR APIs
define an information services API (ISAPI) by which an information service
accesses the other five components (interfaces 13–18, 21, 27, and 33 in Table
14.1(a)). They would also define a CAPI by which the cognition system obtains
status and exerts control over the rest of the system (interfaces 25–30, 5, 11, 17,
23, 29, and 35 in Table 14.1(a)). Although the constituent interfaces of these APIs
are suggested in this table, it would be premature to define these APIs without
first developing detailed information flows and interdependencies. We will define
and analyze these APIs in this chapter. It would also be premature to develop such
APIs without a clear idea of the kinds of RF and user domain knowledge and performance expected of the AACR architecture over time. These aspects are developed in the balance of this chapter, enabling one to draw some conclusions about
these APIs in the final part of this chapter.
A fully defined set of interfaces and APIs would be circumscribed in RXML.
14.2.3 Near-Term Implementations
One way to implement this set of functions is to embed into an SDR a reasoning
engine, such as a rule base with an associated inference engine, as the cognition
function. If the effector functions control parts of the radio, then we have the
simplest AACR based on the simple six-component architecture of Figure 14.2.
Such an approach may be sufficient to expand the control paradigm from today’s
state machines with limited flexibility to tomorrow’s AACR control based
on reasoning over more complex RF states and user situations. Such simple
approaches may well be the next practical steps in AACR evolution from SDR
toward iCR.
This incremental step doesn’t suggest how to mediate the interfaces between
multisensory perception, situation-sensitive prior experience, and a priori knowledge to achieve situation-dependent radio control. Such radio control enables
448
Cognitive Radio Architecture
more sophisticated information services. A simple architecture does not proactively allocate machine-learning (ML) functions to fully understood components.
For example, will AML require an embedded radio propagation modeling tool?
If so, then what is the division of function between a rule base that knows about
radio propagation and a propagation tool that can predict values such as the
received signal-strength indicator (RSSI)? Similarly, in the user domain, some
aspects of user behavior, such as movement by foot and in vehicles, may be modeled in detail based on physics. Will movement modeling be a separate subsystem
based on physics and global positioning system (GPS)? How will that work inside
of buildings? How is the knowledge and skill in tracking user movements divided
between physics-based computational modeling and the symbolic inference of a
rule base or set of Horn clauses4 [34] with a PROLOG engine? For that matter,
how will the learning architecture accommodate a variety of learning methods
such as neural networks, PROLOG, forward chaining, or support vector machines
(SVMs) if learning occurs entirely in a cognition subsystem?
Although hiding such details may be a good thing for AACR in the near term,
it may severely limit the mass customization needed for AACRs to learn user patterns and thus to deliver RF services dramatically better than mere SDRs. Thus,
we need to go further “inside” the cognition and perception subsystems to establish more of a fine-grained architecture. This enables one to structure the data sets
and functions that mediate multisensory domain perception of complex scenes
and related learning technologies that can autonomously adapt to user needs and
preferences. The sequel thus proactively addresses the embedding of ML technology into the radio architecture.
Next, consider the networks. Network-independent SDRs retain multiple personalities in local storage, whereas network-dependent SDRs receive alternate
personalities from a supporting network infrastructure—CWNs. High-end SDRs
both retain alternate personalities locally and have the ability to validate and
accept personalities by download from trusted sources. Whatever architecture
emerges must be consistent with the distribution of RXML knowledge aggregated
in a variety of networks from a tightly coupled CWN to the Internet, with a
degree of <Authority/> and trust reflecting the pragmatics of such different
repositories.
4
A Horn clause is a Boolean expression in which no more than one of the Boolean variables
is positive (not negated). Horn clauses are used in artificial intelligence systems to prove
theorems [9].
449
Chapter 14
Thus, the stage is set for the development of CRA. The following sections
address the cognition cycle, the inference hierarchies, and the SDR architecture
embedded into the CRA.
14.2.4 The Cognition Components
Figure 14.1 shows three computational intelligence aspects of CR:
1. Radio knowledge—RXML:RF
2. User knowledge—RXML:User
3. The capacity to learn
The minimalist architecture of Figure 14.2 and the functional interfaces of
Tables 14.1(a) and 14.1(b) do not assist the radio engineer in structuring knowledge, nor do they assist much in integrating ML into the system. Rather, the finegrained architecture developed in this chapter is derived from the functional
requirements to fully develop these three core capabilities.
Radio Knowledge in the Architecture
Radio knowledge has to be translated from the classroom and engineering teams
into a body of computationally accessible, structured technical knowledge about
radio. RXML is the primary enabler and product of this foray into formalization
of radio knowledge. This text starts a process of RXML definition and development that can be brought to fruition only by industry over time. This process is
similar to the evolution of the SCA of the SDR Forum [3]. The SCA structures
the technical knowledge of the radio components into UML and XML. RXML
will enable the structuring of sufficient RF and user world knowledge to build
advanced wireless-enabled or enhanced information services. Thus, whereas the
SRA and SCA focus on building radios, RXML focuses on using radios.
The World Wide Web (WWW) is now sprouting with computational ontologies some of which are nontechnical but include radio, such as the open Cyc5
5
Cyc is an artificial intelligence (AI) project that attempts to assemble an encyclopedic comprehensive ontology and database of everyday common sense knowledge, with the goal of enabling
AI applications to perform human-like reasoning [9].
450
Cognitive Radio Architecture
ontology. They bring the radio domain into the Semantic Web, which helps
people know about radio. This informal knowledge lacks the technical scope,
precision, and accuracy of authoritative radio references such as the European
Telecommunications Standards Institute (ETSI) documents defining the
Global System for Mobile Communications (GSM) and the International
Telecommunication Union (ITU) definitions of, for example 3GPP.6
Not only must radio knowledge be precise, it must be stated at a useful level
of abstraction, yet with the level of detail appropriate to the use case. Thus, ETSI
GSM in most cases would over-kill the level of detail without providing sufficient
knowledge of the user-centric functionality of GSM. In addition, AACR is multiband, multimode radio (MBMMR), so the knowledge must be comprehensive,
addressing the majority of radio bands and modes available to an MBMMR.
This knowledge is formalized with precision that should be acceptable to ETSI,
the ITU, and regulatory authorities (RAs), yet also be at a level of abstraction
appropriate to internal reasoning, formal dialog with a CWN, or informal dialog
with users.
The capabilities required for an AACR node to be a cognitive entity are to sense,
perceive, orient, plan, decide, act, and learn. To relate ITU standards to these
required capabilities is a process of extracting content from highly formalized
knowledge bases that exist in a unique place and that bear substantial authority,
encapsulating that knowledge in less complete and therefore somewhat approximate
form that can be reasoned with on the AACR node and in real time to support
RF-related use cases. Table 14.2 illustrates this process.
Table 14.2 is illustrative and not comprehensive, but it characterizes the technical issues that drive an information-oriented AACR node architecture. Where
ITU, ETSI, … (meaning other regional and local standards bodies), and CWN
supply source knowledge, the CWN is the repository for authoritative knowledge
derived from the standards bodies and RAs, the <Authorities/>. A user-oriented
AACR may note differences in the interpretation of source knowledge from
<Authorities/> between alternate CWNs, precipitating further knowledge
exchanges.
6
3GPP is a collaboration agreement for the 3G portable phone among ETSI (Europe), the
Association of Radio Industries and Businesses/Telecommunication Technology Committee
(ARIB/TTC; Japan), the China Communications Standards Association (CCSA; China), the Alliance
for Telecommunications Industry Solutions (ATIS; North America), and the Telecommunications
Technology Association (TTA; South Korea).
451
Chapter 14
Table 14.2: Radio knowledge in the node architecture.
Need
Source knowledge
AACR internalization
Sense RF
RF platform
Perceive RF
ITU, ETSI,
ARIB, RAs
Unknown RF
Calibration of RF, noise floor, antennas,
direction
Location-based table of radio spectrum
allocation
RF sensor measurements and knowledge
of basic types (AM, FM, simple digital
channel symbols, typical TDMA, FDMA,
CDMA signal structures)
Receive, parse, and interpret policy
language
Measure parameters in RF, space, and time
Enable SDR for which licensing is current
Optimize transmitted waveform,
space–time plan
Defer spectrum use to legacy users
per policy
Query for available services (white/yellow
pages)
Obtain new skills encapsulated as download
Observe RF
(sense and
perceive)
Orient
XG-like policy
Plan
Known waveform
Known waveform
Restrictive policy
Decide
Act
Legacy waveform,
policy
Applications layer
Learn
ITU, ETSI, …
CWN
Unknown RF
ITU, ETSI, …
CWN
Remember space–time–RF signatures;
discover spectrum use norms and
exceptions
Extract relevant aspects such as new feature
User Knowledge in the Architecture
Next, user knowledge is formalized at the level of abstraction and degree of detail
necessary to give the CR the ability to acquire, from its owner and other designated
users, the user knowledge relevant to information services incrementally. Incremental knowledge acquisition was motivated in the introduction to ML by
describing how frequent occurrences with similar activity sequences identifies
452
Cognitive Radio Architecture
learning opportunities. ML machines may recognize these opportunities for learning through joint probability statistics <Histogram/>. Effective use cases clearly
identify the classes of user and the specific knowledge learned to customize envisioned services. Use cases may also supply sufficient initial knowledge to render
incremental ML not only effective, but also—if possible—enjoyable to the user.
This knowledge is defined in RXML:User. As with RF knowledge, the capabilities required for an AACR node to be a cognitive entity are to OOPDAL. To
relate a use case to these capabilities, one extracts specific and easily recognizable
<Anchors/> for stereotypical situations observable in diverse times, places,
and situations. One expresses the anchor knowledge in RXML for use on the
AACR node.
Cross-domain Grounding for Flexible Information Services
The knowledge about radio and about user needs for wireless services must be
expressed internally in a consistent form so that information services relationships
may be autonomously discovered and maintained by the <Self/> on behalf of the
<User/>. Figure 14.4 shows relationships among user and RF domains.
Staying better connected requires the normalization of knowledge between
<User/> and <RF/> domains. If, for example, the <User/> says, “What’s on one
Observe User
in the Space–Time
Environment
Act for User
in the Space–Time
Environment
4
Observe Radio
in the RF Environment
5
Be
tte
rI
nf
or
m
ed
6 Autonomous
Collaboration
d
te
c
ne
on
C
r
tte
Be
Act for Radio
in the RF Environment
4 Change band/mode to stay connected/ to reduce cost to user/user versus network goal
5 Change data rate, filtering, source, power to optimize type and QoI
6 Manage power, bandwidth, data rate, direction for a community of CR's (CWN, FedNet)
Figure 14.4: Discovering and maintaining services (© Dr. Joseph Mitola III, used with
permission).
453
Chapter 14
oh seven, seven,” near the Washington, DC, area, then the dynamic <User/> ontology should enable the CR to infer that the user is talking about the current frequency modulation (FM) radio broadcast, the units are in MHz, and the user
wants to know what is on WTOP. If it can’t infer this, then it should ask the user
or discover by first dialing a reasonable default, such as 107.7 FM, a broadcast
radio station, and asking, “Is this the radio station you want?” Steps 4, 5, and 6 in
Figure 14.4 all benefit from agreement across domains on how to refer to radio
services. Optimizing behavior to best support the user requires continually adapting the <User/> ontology with repeated regrounding of terms in the <User/>
domain to conceptual primitives and actions in the <RF/> domain.
The CRA facilitates this by seeding the speech recognition subsystem with the
most likely expressions a particular <User/> employs when referring to information services. These would be acquired from the specific users via text and speech
recognition, with dialogs oriented toward continual grounding by posing “yes/no”
questions, either verbally or in displays or both, and obtaining reinforcement
either verbally or via haptic interaction or both. The required degree of mutual
grounding would benefit from specific grounding-oriented features in the AACR
information architecture [12].
The process of linking user expressions of interest to the appropriate radio
technical operations sometimes may be extremely difficult. Military radios, for
example, have many technical parameters: a “channel” in SINCGARS7 consists of
dehopped digital voice in one context (voice communications) or a 25 kHz band
of spectrum in another context (and that may be either an FM channel into which
its frequency hop waveform has hopped or a frequency division multiple access
(FDMA) channel when in single-channel mode). If the user says, “Give me the
commander’s channel,” the SINCGARS user is talking about a “dehopped CVSD
voice stream.”8 If the same user a few seconds later says, “This sounds awful.
Who else is in this channel?” the user is referring to interference with a collection
of hop sets. If the CR observes, “There is strong interference in almost half of
your assigned channels,” then the CR is referring to a related set of 25 kHz channels. If the user then says, “OK, notch the strongest three interference channels,”
the user is talking about a different subset of the channels. If in the next breath the
user says, “Is anything on our emergency channel?” then the user has switched
from SINCGARS context to <Self/> context, asking about one of the cognitive
7
SINCGARS is the Single-Channel Ground and Airborne Radio System used by the US DoD
(see Chapter 4).
8
CVSD stands for continuously variable slope delta modulation, a voice coding method.
454
Cognitive Radio Architecture
military radio’s physical RF access channels. The complexity of such exchanges
demands cross-domain grounding; and the necessity of communicating accurately
but quickly under stress motivates a structured NL and rich radio ontology aspects
of the architecture, developed further in this section.
Thus, both commercial and military information services entail cross-domain
grounding with ontology oriented to NL in the <User/> domain and oriented to
RXML formalized a priori knowledge in the <RF/> domain. Specific methods of
cross-domain grounding with associated architectural features include:
1. <RF/> to <User/> shaping dialog to express precise <RF/> concepts to nonexpert users in an intuitive way, such as:
(a) Grounding: “If you move the speaker box a little bit, it can make a big difference in how well the remote speaker is connected to the wireless transmitter on the television (TV).”
(b) AACR information architecture: Include facility for rich set of synonyms to
mediate cognition–NL–synthesis interface (<Antenna> <Wirelessremote-speaker> “Speaker box”).
2. <RF/> to <User/> learning jargon to express <RF> connectivity opportunities in <User/> terms:
(a) Grounding: “Tee oh pee” for “WTOP,” “Hot ninety-two” for “FM 92.3.”
(b) AACR information architecture: NL-visual facility for single-instance
update of user jargon.
3. <User/> to <RF/> relating values to actions: Relate <User> expression of
values (“low-cost”) to features of situations (“normal”) that are computable
(<NOT> (<CONTAINS> <Situation> <Unusual/></..>)) and that relate
directly to <RF> domain decisions:
(a) Grounding: Normally wait for free wireless local area network (WLAN) for
big attachment; if situation is <Unusual/>, ask if user wants to pay for 3G.
(b) AACR information architecture: Associative inference hierarchy that relates
observable features of a <Scene/> to user sensitivities, such as <Late-forwork/> > <Unusual/>; “The President of the company needs this” >
<Unusual/> because “President” > <VIP/> and <VIP/> is not in most scenes.
14.2.5 Self-referential Components
The cognition component must to assess, manage, and control all of its own
resources, including validating downloads. Thus, in addition to <RF> and <User>
455
Chapter 14
domains, RXML must describe the <Self/>, defining the AACR architecture to the
AACR itself in RXML.
Self-referential Inconsistency
This class of self-referential reasoning is well known in the theory of computing
to be a potential black hole for computational resources. Specifically, any Turingcapable (TC) computational entity that reasons about itself can encounter unexpected Gödel-Turing situations from which it cannot recover. Thus, TC systems
are known to be “partial”—only partially defined because the result obtained
when attempting to execute certain classes of procedure are not definable (the
computing procedure will never terminate).
To avoid this paradox, CR architecture mandates the use of only “total” functions, typically restricted to bounded minimalization [10]. Watchdog “step-counting”
functions [11] or timers must be in place in all its self-referential reasoning and
radio functions. The timer and related computationally indivisible control construct are equivalent to the computer-theoretic construct of a step-counting function over “finite minimalization.” It has been proven that computations that are
limited with certain classes of reliable watchdog timers on finite computing
resources can avoid the Gödel-Turing paradox or at least reduce it to the reliability
of the timer. This proof is the fundamental theorem for practical self-modifying
systems.
In brief, if a system can compute in advance the amount of time or the
number of instructions that any given computation should take, then if that
time or step-count is exceeded, the procedure returns a fixed result such as
“Unreachable in Time T.” As long as the algorithm does not explicitly or
implicitly restart itself on the same problem, then with the associated invocation
of a tightly time- and computationally-constrained alternative tantamount to
giving up, it
(a) is not TC, but
(b) is sufficiently computationally capable to perform real-time communications
tasks such as transmitting and receiving data as well as bounded user interface
functions, and
(c) is not susceptible to the Gödel-Turing incompleteness dilemma, and thus
(d) will not crash because of consuming unbounded or unpredictable resources in
unpredictable self-referential loops.
456
Cognitive Radio Architecture
This is not a general result. This is a highly radio domain-specific result that
has been established only for isochronous communications domains in which
(a) processes are defined in terms of a priori tightly bounded time epochs such as
code division multiple access (CDMA) frames and Signaling System 7 (SS7)
time-outs;9 and
(b) for every situation, there is a default action that has been identified in advance
that consumes O(1) resources; and
(c) the watchdog timer or step-counting function is reliable.
Because radio air interfaces transmit and receive data, there are always defaults
such as “repeat the last packed” or “clear the buffer” that may degrade the performance of the overall communications system. A default has O(1) complexity and
the layers of the protocol stack can implement the default without using
unbounded computing resources.
Watchdog Timer
Without the reliable watchdog timer in the architecture and without this proof to
establish the rules for acceptable computing constructs on CRs, engineers and computer programmers would build CRs that would crash in extremely unpredictable
ways as their adaptation algorithms got trapped in unpredictable unbounded self-referential loops. Because planning problems exist that cannot be solved with algorithms so constrained, either an unbounded community of CRs must cooperatively
work on the more general problems or the cognitive network (CN) must employ a
TC algorithm to solve the more difficult problems (e.g., NP-hard with large N )
offline. There is also the interesting possibility of trading off space and time by
remembering partial solutions and restarting NP-hard problems with these subproblems already solved. Although it doesn’t actually avoid any necessary calculations, with O(N) pattern matching for solved subproblems, it may reduce the
total computational burden, somewhat.10 This class of approach to parallel problem-solving is similar to the use of pheromones by ants to solve the traveling
9
SS7 is the seventh and most stable software update of the long-distance telephony software.
Time-out here means that the software expected tasks to complete by a certain time-out; if not, it
determined that something had gone awry, and either tried a different route or delivered a tone
indicating that a system fault had occurred.
10
For example, the fast Fourier transform (FFT) converts O(N2) steps to O(N log N) by avoiding
the recomputation of already computed partial products.
457
Chapter 14
salesman problem in less than (2N)/M time with M ants. But this is an engineering
text, not a text on the theory of computing, so these aspects are not developed further here, but it suffices to show the predictable finiteness and proof that the
approach is boundable and hence compatible with the real-time performance
needs of CR.
This timer-based finite computing regime also works for user interfaces
because users will not wait forever before changing the situation (e.g., by shutting
off the radio or hitting another key); and the CR can always kind of throw up its
hands and ask the user to take over.
Thus, with a proof of stability based on the theory of computing, the CRA
structures systems that not only can mo